{ "version": 2, "date": "2026-11-08", "title": "模型推理状态：整段复用与 checkpoint 分账", "modelCount": 109, "originalModelCount": 50, "newModelCount": 59, "defaults": { "tokens": 32768, "interval": 512, "kvBits": 16, "stateBits": 32, "windowMode": "dedup" }, "models": [ { "id": "DeepSeek-V4.1-Flash", "repo": "deepseek-ai/DeepSeek-V4.1-Flash", "name": "DeepSeek V4.1 Flash", "family": "deepseek-ai", "params": "552B + 196B Engram / A8–16B", "size": 552, "arch": "共享压缩 latent + SWA", "tier": "现役代表", "layers": 40, "maxContext": 1048576, "components": [ { "label": "40 层最近窗口 latent", "elements": 20480, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" }, { "label": "1 组共享 1:1 latent", "elements": 512, "kind": "compressed", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "1 组共享 1:1 index K", "elements": 128, "kind": "compressed", "ratio": 1, "window": null, "role": "index", "dtype": "selected" }, { "label": "3 组共享 2:1 latent", "elements": 1536, "kind": "compressed", "ratio": 2, "window": null, "role": "main", "dtype": "selected" }, { "label": "3 组共享 2:1 index K", "elements": 384, "kind": "compressed", "ratio": 2, "window": null, "role": "index", "dtype": "selected" }, { "label": "FP32 压缩尾组状态", "elements": 6144, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "fp32", "stateClass": "compressor" }, { "label": "Engram 压缩 token ID / DEAD 历史", "elements": 3, "kind": "fixed", "dtype": "int64", "role": "state", "ratio": 1, "window": null, "stateClass": "ngram" } ], "notes": [ "只有层 2、8、14、20 拥有独立全局主 KV 和 index K；压缩比为 2、2、2、1。其余层复用，所有 40 层各自保留 128-token 局部窗口。", "模型卡另报原生混合量化全局缓存 890 B/token（FP4 主 KV，含索引等）；它不含本表的局部窗口项，不与统一 BF16 排名混用。", "本页补入 Engram 最后 3 个 int64 压缩 token ID（含 DEAD 标记）。官方参考实现预分配整段 ID 数组；逻辑恢复载荷只需这 3 个，静态词表 / hash 参数属于权重。" ], "extraSources": [ { "title": "V4.1 CED/CSA2 共享缓存实现", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/inference/model.py" }, { "title": "Engram NgramHashState 的有限 look-back", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/inference/engram.py" } ], "configSha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash", "evidence": "配置 + 共享缓存实现核验", "geometry": { "windowLayers": 40, "window": 128, "latentDim": 512, "indexDim": 128, "globalSources": [ 2, 8, 14, 20 ], "sourceRatios": [ 2, 2, 2, 1 ] }, "formula": "40 × 512 × b × min(T,128) + (3floor(T/2)+T) × (512+128) × b + 3 × 2 × 2 × 512 × 4", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 110125080.0, "arbitrary": 2253127680.0, "checkpoint": 441976320.0, "checkpointIndependent": 441976320.0, "globalBytes": 104857600.0, "windowBytes": 5242880.0, "stateBytes": 24600, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 24600, "checkpointCount": 64, "growthBytesPerToken": 3200.0, "endpointBytesPerToken": 3360.750732421875, "checkpointBytesPerToken": 13488.046875, "checkpointIndependentBytesPerToken": 13488.046875, "arbitraryBytesPerToken": 68760.0 }, "defaultRanks": { "endpoint": 2, "checkpoint": 3, "checkpointIndependent": 3, "arbitrary": 12, "growthBytesPerToken": 3, "stateBytes": 66 } }, { "id": "gemma-4-E2B", "repo": "google/gemma-4-E2B", "name": "Gemma 4 E2B", "family": "google", "params": "5.1B / 有效 2.3B", "size": 5.1, "arch": "GQA + SWA", "tier": "现役代表", "layers": 35, "maxContext": 131072, "components": [ { "label": "全局 K + V（独立缓存层）", "elements": 3072, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "滑窗 K + V（独立缓存层）", "elements": 6144, "kind": "window", "ratio": 1, "window": 512, "role": "main", "dtype": "selected" } ], "notes": [ "20 个消费者层复用已有 K/V，只对 15 个独立缓存层计数。", "SWA 按执行窗口回收后的驻留量计算；保留所有历史滑窗 KV 的后端会占用更多。" ], "extraSources": [ { "title": "Gemma 4 共享层映射", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/configuration_gemma4.py" }, { "title": "Gemma 4 K/V 缓存实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py" } ], "configSha256": "e5faef0dd1a8f2437f6010721146b85433eaa90e679ef011e803c7ffefae73b8", "configUrl": "https://huggingface.co/google/gemma-4-E2B/blob/main/config.json", "cardUrl": "https://huggingface.co/google/gemma-4-E2B", "evidence": "配置 + 共享语义核验", "geometry": { "fullLayers": 3, "windowLayers": 12, "sharedLayers": 20, "kvHeads": 1, "keyDim": 512, "valueDim": 512, "localKvHeads": 1, "localHeadDim": 256, "window": 512 }, "formula": "3 × 2 × 1 × 512 × b × T + 12 × 2 × 1 × 256 × b × min(T,512)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 207618048.0, "arbitrary": 603979776.0, "checkpoint": 603979776.0, "checkpointIndependent": 603979776.0, "globalBytes": 201326592.0, "windowBytes": 6291456.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 6144.0, "endpointBytesPerToken": 6336.0, "checkpointBytesPerToken": 18432.0, "checkpointIndependentBytesPerToken": 18432.0, "arbitraryBytesPerToken": 18432.0 }, "defaultRanks": { "endpoint": 4, "checkpoint": 5, "checkpointIndependent": 5, "arbitrary": 1, "growthBytesPerToken": 4, "stateBytes": 1 } }, { "id": "DeepSeek-V4-Flash", "repo": "deepseek-ai/DeepSeek-V4-Flash", "name": "DeepSeek V4-Flash", "family": "deepseek-ai", "params": "284B / A13B", "size": 284, "arch": "压缩 latent + SWA", "tier": "现役代表", "layers": 43, "maxContext": 1048576, "components": [ { "label": "最近窗口 latent", "elements": 22016, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" }, { "label": "4:1 主 latent", "elements": 10752, "kind": "compressed", "ratio": 4, "window": null, "role": "main", "dtype": "selected" }, { "label": "4:1 index K", "elements": 2688, "kind": "compressed", "ratio": 4, "window": null, "role": "index", "dtype": "selected" }, { "label": "128:1 主 latent", "elements": 10240, "kind": "compressed", "ratio": 128, "window": null, "role": "main", "dtype": "selected" }, { "label": "FP32 压缩器 kv/score 状态", "elements": 3051520, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "fp32", "stateClass": "compressor" } ], "notes": [ "主 latent 由 K/V 共用，512 维已经含 64 维 RoPE 分支；不再加 64、不再乘 2。", "compress_ratios 末项是 MTP 层，已排除；4:1 重叠压缩器与 128:1 压缩器的 FP32 持久状态按官方参考代码计算。", "这里只比较统一 BF16/8-bit 载荷；官方混合 KV 量化、缓存布局和高效压缩器实现可能改变实际占用。" ], "extraSources": [ { "title": "DeepSeek V4 压缩器及缓存实现", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/inference/model.py" } ], "configSha256": "b628e63398a645abc711d92207f8737dd8140f7a4ef1e0a5b3616019e0ddd818", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash", "evidence": "配置 + 压缩缓存实现核验", "geometry": { "windowLayers": 43, "window": 128, "latentDim": 512, "indexDim": 128, "ratioCounts": { "0": 2, "4": 21, "128": 20 } }, "formula": "43 × 512 × b × min(T,128) + Σ[L_r × 512 × b × floor(T/r)] + 21 × 128 × b × floor(T/4) + S_compressor", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 243286016.0, "arbitrary": 401637113856.0, "checkpoint": 1367343104.0, "checkpointIndependent": 1367343104.0, "globalBytes": 225443840.0, "windowBytes": 5636096.0, "stateBytes": 12206080, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 12206080, "checkpointCount": 64, "growthBytesPerToken": 6880.0, "endpointBytesPerToken": 7424.5, "checkpointBytesPerToken": 41728.0, "checkpointIndependentBytesPerToken": 41728.0, "arbitraryBytesPerToken": 12256992.0 }, "defaultRanks": { "endpoint": 5, "checkpoint": 11, "checkpointIndependent": 11, "arbitrary": 70, "growthBytesPerToken": 6, "stateBytes": 70 } }, { "id": "DeepSeek-V4-Pro", "repo": "deepseek-ai/DeepSeek-V4-Pro", "name": "DeepSeek V4-Pro", "family": "deepseek-ai", "params": "1.6T / A49B", "size": 1600, "arch": "压缩 latent + SWA", "tier": "现役代表", "layers": 61, "maxContext": 1048576, "components": [ { "label": "最近窗口 latent", "elements": 31232, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" }, { "label": "128:1 主 latent", "elements": 15872, "kind": "compressed", "ratio": 128, "window": null, "role": "main", "dtype": "selected" }, { "label": "4:1 主 latent", "elements": 15360, "kind": "compressed", "ratio": 4, "window": null, "role": "main", "dtype": "selected" }, { "label": "4:1 index K", "elements": 3840, "kind": "compressed", "ratio": 4, "window": null, "role": "index", "dtype": "selected" }, { "label": "FP32 压缩器 kv/score 状态", "elements": 4677632, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "fp32", "stateClass": "compressor" } ], "notes": [ "主 latent 由 K/V 共用，512 维已经含 64 维 RoPE 分支；不再加 64、不再乘 2。", "compress_ratios 末项是 MTP 层，已排除；4:1 重叠压缩器与 128:1 压缩器的 FP32 持久状态按官方参考代码计算。", "这里只比较统一 BF16/8-bit 载荷；官方混合 KV 量化、缓存布局和高效压缩器实现可能改变实际占用。" ], "extraSources": [ { "title": "DeepSeek V4 压缩器及缓存实现", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/inference/model.py" } ], "configSha256": "5fe4568daee51c208cb8a79538eaeda090ae011ade1dee2c386aa95f569c810e", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro", "evidence": "配置 + 压缩缓存实现核验", "geometry": { "windowLayers": 61, "window": 128, "latentDim": 512, "indexDim": 128, "ratioCounts": { "128": 31, "4": 30 } }, "formula": "61 × 512 × b × min(T,128) + Σ[L_r × 512 × b × floor(T/r)] + 30 × 128 × b × floor(T/4) + S_compressor", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 349405184.0, "arbitrary": 615476101120.0, "checkpoint": 2031878144.0, "checkpointIndependent": 2031878144.0, "globalBytes": 322699264.0, "windowBytes": 7995392.0, "stateBytes": 18710528, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 18710528, "checkpointCount": 64, "growthBytesPerToken": 9848.0, "endpointBytesPerToken": 10663.0, "checkpointBytesPerToken": 62008.0, "checkpointIndependentBytesPerToken": 62008.0, "arbitraryBytesPerToken": 18782840.0 }, "defaultRanks": { "endpoint": 11, "checkpoint": 19, "checkpointIndependent": 19, "arbitrary": 71, "growthBytesPerToken": 12, "stateBytes": 71 } }, { "id": "GLM-5.3-Flash", "repo": "zai-org/GLM-5.3-Flash", "name": "GLM-5.3 Flash", "family": "zai-org", "params": "320B / A18B", "size": 320, "arch": "MLA + KDA + DSA", "tier": "现役代表", "layers": 45, "maxContext": 1048576, "components": [ { "label": "NoPE MLA latent", "elements": 5632, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "4:1 池化 index K", "elements": 1408, "kind": "compressed", "ratio": 4, "window": null, "role": "index", "dtype": "selected" }, { "label": "KDA 递归状态", "elements": 35651584, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "KDA BF16 短卷积历史", "elements": 2506752, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" }, { "label": "DSA 池化 key + gate-score 尾块", "elements": 11264, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "index-tail" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。", "本页补全 index_kpool=4 的 key 和 gate-score 两份 BF16 尾块；不启用投机解码 extra slots。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "池化 DSA indexer", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/layers/attention/dsa/dsa_indexer_kpool.py" }, { "title": "DSA BF16 双尾块（不含投机 extra slots）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/mem_cache/memory_pool.py" } ], "configSha256": "bb8f01c42cb92a52ca72e65afb4d5bd8d11aef083cd210e8de25dfb904f23e9f", "configUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-5.3-Flash", "evidence": "配置 + 公式", "geometry": { "fullLayers": 11, "latentDim": 512, "positionalDim": 0, "indexLayers": 11, "indexDim": 128, "indexRatio": 4, "linearLayers": 34, "stateHeads": 64, "stateKeyHeads": 64, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "11 × 512 × b × T + 11 × 128 × b × floor(T/4) + S_KDA", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 539809792.0, "arbitrary": 4838337282048.0, "checkpoint": 9841278976.0, "checkpointIndependent": 9841278976.0, "globalBytes": 392167424.0, "windowBytes": 0, "stateBytes": 147642368.0, "recurrentBytes": 142606336.0, "convBytes": 5013504, "otherStateBytes": 22528.0, "checkpointCount": 64, "growthBytesPerToken": 11968.0, "endpointBytesPerToken": 16473.6875, "checkpointBytesPerToken": 300332.0, "checkpointIndependentBytesPerToken": 300332.0, "arbitraryBytesPerToken": 147654336.0 }, "defaultRanks": { "endpoint": 17, "checkpoint": 80, "checkpointIndependent": 78, "arbitrary": 90, "growthBytesPerToken": 13, "stateBytes": 90 } }, { "id": "Qwen3.5-2B", "repo": "Qwen/Qwen3.5-2B", "name": "Qwen3.5 2B", "family": "Qwen", "params": "2B", "size": 2, "arch": "GQA + GDN", "tier": "现役代表", "layers": 24, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 6144, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 4718592, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 331776, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "ed1c1723241f23f7f4e23430759cbd7dcfb4103cbdfe052bfe7626b57c2615b4", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-2B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-2B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 6, "kvHeads": 2, "keyDim": 256, "valueDim": 256, "linearLayers": 18, "stateHeads": 16, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "6 × 2 × 2 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 422191104.0, "arbitrary": 640621215744.0, "checkpoint": 1653080064.0, "checkpointIndependent": 1653080064.0, "globalBytes": 402653184.0, "windowBytes": 0, "stateBytes": 19537920.0, "recurrentBytes": 18874368.0, "convBytes": 663552, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 12288.0, "endpointBytesPerToken": 12884.25, "checkpointBytesPerToken": 50448.0, "checkpointIndependentBytesPerToken": 50448.0, "arbitraryBytesPerToken": 19550208.0 }, "defaultRanks": { "endpoint": 14, "checkpoint": 14, "checkpointIndependent": 14, "arbitrary": 72, "growthBytesPerToken": 14, "stateBytes": 72 } }, { "id": "gemma-4-E4B", "repo": "google/gemma-4-E4B", "name": "Gemma 4 E4B", "family": "google", "params": "8B / 有效 4.5B", "size": 8, "arch": "GQA + SWA", "tier": "现役代表", "layers": 42, "maxContext": 131072, "components": [ { "label": "全局 K + V（独立缓存层）", "elements": 8192, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "滑窗 K + V（独立缓存层）", "elements": 20480, "kind": "window", "ratio": 1, "window": 512, "role": "main", "dtype": "selected" } ], "notes": [ "18 个消费者层复用已有 K/V，只对 24 个独立缓存层计数。", "SWA 按执行窗口回收后的驻留量计算；保留所有历史滑窗 KV 的后端会占用更多。" ], "extraSources": [ { "title": "Gemma 4 共享层映射", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/configuration_gemma4.py" }, { "title": "Gemma 4 K/V 缓存实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py" } ], "configSha256": "f27a045f32c39fb9cd930204920de6b0962810cf09929a8810901ffaec780f20", "configUrl": "https://huggingface.co/google/gemma-4-E4B/blob/main/config.json", "cardUrl": "https://huggingface.co/google/gemma-4-E4B", "evidence": "配置 + 共享语义核验", "geometry": { "fullLayers": 4, "windowLayers": 20, "sharedLayers": 18, "kvHeads": 2, "keyDim": 512, "valueDim": 512, "localKvHeads": 2, "localHeadDim": 256, "window": 512 }, "formula": "4 × 2 × 2 × 512 × b × T + 20 × 2 × 2 × 256 × b × min(T,512)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 557842432.0, "arbitrary": 1879048192.0, "checkpoint": 1879048192.0, "checkpointIndependent": 1879048192.0, "globalBytes": 536870912.0, "windowBytes": 20971520.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 16384.0, "endpointBytesPerToken": 17024.0, "checkpointBytesPerToken": 57344.0, "checkpointIndependentBytesPerToken": 57344.0, "arbitraryBytesPerToken": 57344.0 }, "defaultRanks": { "endpoint": 18, "checkpoint": 16, "checkpointIndependent": 16, "arbitrary": 9, "growthBytesPerToken": 17, "stateBytes": 1 } }, { "id": "gemma-4-12B", "repo": "google/gemma-4-12B", "name": "Gemma 4 12B", "family": "google", "params": "11.95B", "size": 11.95, "arch": "GQA + SWA", "tier": "现役代表", "layers": 48, "maxContext": 262144, "components": [ { "label": "全局 K + V（独立缓存层）", "elements": 8192, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "滑窗 K + V（独立缓存层）", "elements": 163840, "kind": "window", "ratio": 1, "window": 1024, "role": "main", "dtype": "selected" } ], "notes": [ "attention_k_eq_v 是投影权重共享；K 经过 RoPE、V 不经过 RoPE，缓存仍分别保存 K 与 V，不能据此再减半。", "SWA 按执行窗口回收后的驻留量计算；保留所有历史滑窗 KV 的后端会占用更多。" ], "extraSources": [ { "title": "Gemma 4 共享层映射", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/configuration_gemma4.py" }, { "title": "Gemma 4 K/V 缓存实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py" } ], "configSha256": "14f38c5492ffc9cbcdf808647ca0c025bb5b9b4eb737526347134d500ace6098", "configUrl": "https://huggingface.co/google/gemma-4-12B/blob/main/config.json", "cardUrl": "https://huggingface.co/google/gemma-4-12B", "evidence": "配置 + 共享语义核验", "geometry": { "fullLayers": 8, "windowLayers": 40, "sharedLayers": 0, "kvHeads": 1, "keyDim": 512, "valueDim": 512, "localKvHeads": 8, "localHeadDim": 256, "window": 1024 }, "formula": "8 × 2 × 1 × 512 × b × T + 40 × 2 × 8 × 256 × b × min(T,1024)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 872415232.0, "arbitrary": 11274289152.0, "checkpoint": 11274289152.0, "checkpointIndependent": 21843935232.0, "globalBytes": 536870912.0, "windowBytes": 335544320.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 16384.0, "endpointBytesPerToken": 26624.0, "checkpointBytesPerToken": 344064.0, "checkpointIndependentBytesPerToken": 666624.0, "arbitraryBytesPerToken": 344064.0 }, "defaultRanks": { "endpoint": 25, "checkpoint": 88, "checkpointIndependent": 94, "arbitrary": 62, "growthBytesPerToken": 17, "stateBytes": 1 } }, { "id": "gemma-4-26B-A4B", "repo": "google/gemma-4-26B-A4B", "name": "Gemma 4 26B-A4B", "family": "google", "params": "25.2B / A3.8B", "size": 25.2, "arch": "GQA + SWA", "tier": "现役代表", "layers": 30, "maxContext": 262144, "components": [ { "label": "全局 K + V（独立缓存层）", "elements": 10240, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "滑窗 K + V（独立缓存层）", "elements": 102400, "kind": "window", "ratio": 1, "window": 1024, "role": "main", "dtype": "selected" } ], "notes": [ "attention_k_eq_v 是投影权重共享；K 经过 RoPE、V 不经过 RoPE，缓存仍分别保存 K 与 V，不能据此再减半。", "SWA 按执行窗口回收后的驻留量计算；保留所有历史滑窗 KV 的后端会占用更多。" ], "extraSources": [ { "title": "Gemma 4 共享层映射", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/configuration_gemma4.py" }, { "title": "Gemma 4 K/V 缓存实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py" } ], "configSha256": "0364256c5de7863a22b5dd666043d542d42b681c02fd83eb67dc7b483d859e96", "configUrl": "https://huggingface.co/google/gemma-4-26B-A4B/blob/main/config.json", "cardUrl": "https://huggingface.co/google/gemma-4-26B-A4B", "evidence": "配置 + 共享语义核验", "geometry": { "fullLayers": 5, "windowLayers": 25, "sharedLayers": 0, "kvHeads": 2, "keyDim": 512, "valueDim": 512, "localKvHeads": 8, "localHeadDim": 256, "window": 1024 }, "formula": "5 × 2 × 2 × 512 × b × T + 25 × 2 × 8 × 256 × b × min(T,1024)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 880803840.0, "arbitrary": 7381975040.0, "checkpoint": 7381975040.0, "checkpointIndependent": 13988003840.0, "globalBytes": 671088640.0, "windowBytes": 209715200.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 20480.0, "endpointBytesPerToken": 26880.0, "checkpointBytesPerToken": 225280.0, "checkpointIndependentBytesPerToken": 426880.0, "arbitraryBytesPerToken": 225280.0 }, "defaultRanks": { "endpoint": 26, "checkpoint": 70, "checkpointIndependent": 92, "arbitrary": 52, "growthBytesPerToken": 22, "stateBytes": 1 } }, { "id": "Qwen3.5-35B-A3B", "repo": "Qwen/Qwen3.5-35B-A3B", "name": "Qwen3.5 35B-A3B", "family": "Qwen", "params": "35B / A3B", "size": 35, "arch": "GQA + GDN", "tier": "现役代表", "layers": 40, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 10240, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 15728640, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 737280, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "5e4d7f74fec2f360eb9cfbfcd6ec0c4c76e684d3a11caaed259d9fd9bfbc7944", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-35B-A3B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 10, "kvHeads": 2, "keyDim": 256, "valueDim": 256, "linearLayers": 30, "stateHeads": 32, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "10 × 2 × 2 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 735477760.0, "arbitrary": 2110573772800.0, "checkpoint": 4791992320.0, "checkpointIndependent": 4791992320.0, "globalBytes": 671088640.0, "windowBytes": 0, "stateBytes": 64389120.0, "recurrentBytes": 62914560.0, "convBytes": 1474560, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 20480.0, "endpointBytesPerToken": 22445.0, "checkpointBytesPerToken": 146240.0, "checkpointIndependentBytesPerToken": 146240.0, "arbitraryBytesPerToken": 64409600.0 }, "defaultRanks": { "endpoint": 21, "checkpoint": 50, "checkpointIndependent": 50, "arbitrary": 79, "growthBytesPerToken": 22, "stateBytes": 79 } }, { "id": "gpt-oss-20b", "repo": "openai/gpt-oss-20b", "name": "gpt-oss-20b", "family": "openai", "params": "约 21B / A3.6B", "size": 21, "arch": "GQA + SWA", "tier": "现役代表", "layers": 24, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "SWA K + V", "elements": 12288, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" } ], "notes": [ "MXFP4 是发布权重的精度，不代表 KV 是 4-bit；此处独立选择 KV 精度。" ], "extraSources": [], "configSha256": "3a2a26ded679375b7928ddeca59764df7cea83220c1961035f6d6e232659e9ce", "configUrl": "https://huggingface.co/openai/gpt-oss-20b/blob/main/config.json", "cardUrl": "https://huggingface.co/openai/gpt-oss-20b", "evidence": "配置 + 公式", "geometry": { "fullLayers": 12, "windowLayers": 12, "kvHeads": 8, "keyDim": 64, "valueDim": 64, "window": 128 }, "formula": "2 × 8 × 64 × b × (12T + 12min(T,128))", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 808452096.0, "arbitrary": 1610612736.0, "checkpoint": 1006632960.0, "checkpointIndependent": 1006632960.0, "globalBytes": 805306368.0, "windowBytes": 3145728.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 24576.0, "endpointBytesPerToken": 24672.0, "checkpointBytesPerToken": 30720.0, "checkpointIndependentBytesPerToken": 30720.0, "arbitraryBytesPerToken": 49152.0 }, "defaultRanks": { "endpoint": 24, "checkpoint": 7, "checkpointIndependent": 7, "arbitrary": 6, "growthBytesPerToken": 26, "stateBytes": 1 } }, { "id": "Qwen3.5-122B-A10B", "repo": "Qwen/Qwen3.5-122B-A10B", "name": "Qwen3.5 122B-A10B", "family": "Qwen", "params": "122B / A10B", "size": 122, "arch": "GQA + GDN", "tier": "现役代表", "layers": 48, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 37748736, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 1327104, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "af07d0423658865afb52bd0d6fd3c2cc45d988425b76193467fd19129655cecd", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-122B-A10B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 12, "kvHeads": 2, "keyDim": 256, "valueDim": 256, "linearLayers": 36, "stateHeads": 64, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "12 × 2 × 2 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 958955520.0, "arbitrary": 5035580719104.0, "checkpoint": 10638852096.0, "checkpointIndependent": 10638852096.0, "globalBytes": 805306368.0, "windowBytes": 0, "stateBytes": 153649152.0, "recurrentBytes": 150994944.0, "convBytes": 2654208, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 24576.0, "endpointBytesPerToken": 29265.0, "checkpointBytesPerToken": 324672.0, "checkpointIndependentBytesPerToken": 324672.0, "arbitraryBytesPerToken": 153673728.0 }, "defaultRanks": { "endpoint": 32, "checkpoint": 83, "checkpointIndependent": 81, "arbitrary": 92, "growthBytesPerToken": 26, "stateBytes": 92 } }, { "id": "Qwen3.8-Flash-Next", "repo": "Qwen/Qwen3.8-Flash-Next", "name": "Qwen3.8 Flash-Next", "family": "Qwen", "params": "125B + 51B n-gram / A6B", "size": 125, "arch": "GQA + GDN + QSA", "tier": "现役代表", "layers": 48, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 28311552, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 1105920, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" }, { "label": "QSA 4:1 池化索引", "elements": 1536, "kind": "compressed", "ratio": 4, "window": null, "role": "index", "dtype": "selected" }, { "label": "QSA pending-key ring（4 槽）", "elements": 6144, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "index-tail" }, { "label": "QSA 共用 MRoPE 坐标 ring", "elements": 12, "kind": "fixed", "dtype": "int64", "role": "state", "ratio": 1, "window": null, "stateClass": "position" }, { "label": "PLE 膨胀卷积历史", "elements": 92160, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" }, { "label": "PLE n-gram token 历史", "elements": 2, "kind": "fixed", "dtype": "int64", "role": "state", "ratio": 1, "window": null, "stateClass": "ngram" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。", "本页补全 QSA key ring、共享的三轴 MRoPE 坐标、PLE 9 槽膨胀卷积历史和 2 个 n-gram token ID。ring 按完整容量快照保存，即使在整组边界部分槽位已无效。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" }, { "title": "QSA 索引与 4:1 池化", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/layers/attention/qsa/qsa_indexer.py" }, { "title": "QSA key / MRoPE ring 的真实形状", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/mem_cache/qsa_kv_pool.py" }, { "title": "PLE 膨胀卷积与 n-gram", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/qwen4_exp.py" } ], "configSha256": "889658f2508e8c61d409b02e70e0d78d8d4452ec65aaafbe129805d213d2e74b", "configUrl": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.8-Flash-Next", "evidence": "配置 + 公式", "geometry": { "fullLayers": 12, "kvHeads": 2, "keyDim": 256, "valueDim": 256, "linearLayers": 36, "stateHeads": 48, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "12 × 2 × 2 × 256 × b × T + S_GDN + 12 × 128 × b × floor(T/4)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 946126960.0, "arbitrary": 3790605910016.0, "checkpoint": 8232377344.0, "checkpointIndependent": 8232377344.0, "globalBytes": 830472192.0, "windowBytes": 0, "stateBytes": 115654768.0, "recurrentBytes": 113246208.0, "convBytes": 2396160, "otherStateBytes": 12400.0, "checkpointCount": 64, "growthBytesPerToken": 25344.0, "endpointBytesPerToken": 28873.50341796875, "checkpointBytesPerToken": 251232.21875, "checkpointIndependentBytesPerToken": 251232.21875, "arbitraryBytesPerToken": 115680112.0 }, "defaultRanks": { "endpoint": 31, "checkpoint": 72, "checkpointIndependent": 71, "arbitrary": 87, "growthBytesPerToken": 31, "stateBytes": 87 } }, { "id": "Kimi-K3", "repo": "moonshotai/Kimi-K3", "name": "Kimi K3", "family": "moonshotai", "params": "2.8T / A104B", "size": 2800, "arch": "MLA + KDA", "tier": "现役代表", "layers": 93, "maxContext": 1048576, "components": [ { "label": "24 层 MLA latent + 64 维分支", "elements": 13824, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "KDA 递归状态", "elements": 108527616, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "KDA BF16 短卷积历史", "elements": 7630848, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。", "mla_use_nope=true 不代表 64 维分支消失：官方投影与拼接仍保留该分支，因此是 576 维压缩载荷，不是 512 维。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Kimi K3 模型代码", "url": "https://huggingface.co/moonshotai/Kimi-K3/blob/main/modeling_kimi_linear.py" }, { "title": "Kimi K3 技术报告", "url": "https://arxiv.org/html/2607.24653v1" } ], "configSha256": "9710e121a58d03ac92c8d6da287a19541994319afbbe6d6202af001ffd379213", "configUrl": "https://huggingface.co/moonshotai/Kimi-K3/blob/main/config.json", "cardUrl": "https://huggingface.co/moonshotai/Kimi-K3", "evidence": "配置 + 模型代码 + 报告", "geometry": { "fullLayers": 24, "latentDim": 512, "positionalDim": 64, "linearLayers": 69, "stateHeads": 96, "stateKeyHeads": 96, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "24 × (512+64) × b × T + 69 × 96 × 128 × 128 × b_state + 69 × 3 × (3×96×128) × 2", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1355341824.0, "arbitrary": 14725932908544.0, "checkpoint": 29665787904.0, "checkpointIndependent": 29665787904.0, "globalBytes": 905969664.0, "windowBytes": 0, "stateBytes": 449372160.0, "recurrentBytes": 434110464.0, "convBytes": 15261696, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 27648.0, "endpointBytesPerToken": 41361.75, "checkpointBytesPerToken": 905328.0, "checkpointIndependentBytesPerToken": 905328.0, "arbitraryBytesPerToken": 449399808.0 }, "defaultRanks": { "endpoint": 41, "checkpoint": 99, "checkpointIndependent": 96, "arbitrary": 99, "growthBytesPerToken": 32, "stateBytes": 99 } }, { "id": "Qwen2.5-1.5B-Instruct", "repo": "Qwen/Qwen2.5-1.5B-Instruct", "name": "Qwen2.5 1.5B", "family": "Qwen", "params": "1.5B", "size": 1.5, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 28, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 14336, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "配置中的滑窗大于或等于本 checkpoint 的最大上下文；在有效范围内按全长 KV 计算。" ], "extraSources": [], "configSha256": "98d2ff8cc47488d08a2b0b3acf4eb99ef210779b42bd48605f6b8e36acdbf670", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct", "evidence": "配置 + 公式", "geometry": { "fullLayers": 28, "kvHeads": 2, "keyDim": 128, "valueDim": 128, "queryHeads": 12 }, "formula": "2 × 28 层 × 2 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 939524096.0, "arbitrary": 939524096.0, "checkpoint": 939524096.0, "checkpointIndependent": 939524096.0, "globalBytes": 939524096.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 28672.0, "endpointBytesPerToken": 28672.0, "checkpointBytesPerToken": 28672.0, "checkpointIndependentBytesPerToken": 28672.0, "arbitraryBytesPerToken": 28672.0 }, "defaultRanks": { "endpoint": 30, "checkpoint": 6, "checkpointIndependent": 6, "arbitrary": 2, "growthBytesPerToken": 33, "stateBytes": 1 } }, { "id": "Qwen3.5-397B-A17B", "repo": "Qwen/Qwen3.5-397B-A17B", "name": "Qwen3.5 397B-A17B", "family": "Qwen", "params": "397B / A17B", "size": 397, "arch": "GQA + GDN", "tier": "现役代表", "layers": 60, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 15360, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 47185920, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 1658880, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "3ae7fa89c2f7d1354096418ddaf1331e9e0898a8ca11e804fb6dbaa087efb7da", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-397B-A17B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 15, "kvHeads": 2, "keyDim": 256, "valueDim": 256, "linearLayers": 45, "stateHeads": 64, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "15 × 2 × 2 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1198694400.0, "arbitrary": 6294475898880.0, "checkpoint": 13298565120.0, "checkpointIndependent": 13298565120.0, "globalBytes": 1006632960.0, "windowBytes": 0, "stateBytes": 192061440.0, "recurrentBytes": 188743680.0, "convBytes": 3317760, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 30720.0, "endpointBytesPerToken": 36581.25, "checkpointBytesPerToken": 405840.0, "checkpointIndependentBytesPerToken": 405840.0, "arbitraryBytesPerToken": 192092160.0 }, "defaultRanks": { "endpoint": 37, "checkpoint": 94, "checkpointIndependent": 91, "arbitrary": 97, "growthBytesPerToken": 34, "stateBytes": 97 } }, { "id": "DeepSeek-V2-Lite", "repo": "deepseek-ai/DeepSeek-V2-Lite", "name": "DeepSeek V2 Lite", "family": "deepseek-ai", "params": "16B / A2.4B", "size": 16, "arch": "MLA", "tier": "经典 / 广泛使用", "layers": 27, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 15552, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "采用压缩 MLA 缓存；16B 是总参数量，A2.4B 是激活参数量。" ], "extraSources": [], "configSha256": "f346286b0f1c8b044252fd54cb4fa78b9fab6472a6e8bebb9edfe03d414ea03d", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite", "evidence": "配置 + 公式", "geometry": { "fullLayers": 27, "latentDim": 512, "positionalDim": 64 }, "formula": "27 × (512+64) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1019215872.0, "arbitrary": 1019215872.0, "checkpoint": 1019215872.0, "checkpointIndependent": 1019215872.0, "globalBytes": 1019215872.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 31104.0, "endpointBytesPerToken": 31104.0, "checkpointBytesPerToken": 31104.0, "checkpointIndependentBytesPerToken": 31104.0, "arbitraryBytesPerToken": 31104.0 }, "defaultRanks": { "endpoint": 33, "checkpoint": 8, "checkpointIndependent": 8, "arbitrary": 3, "growthBytesPerToken": 35, "stateBytes": 1 } }, { "id": "Llama-3.2-1B-Instruct", "repo": "meta-llama/Llama-3.2-1B-Instruct", "name": "Llama 3.2 1B", "family": "meta-llama", "params": "1B", "size": 1, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 16, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 16384, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "geometry": { "fullLayers": 16, "kvHeads": 8, "keyDim": 64, "valueDim": 64 }, "formula": "2 × 16 层 × 8 KV heads × 64 dim × b", "notes": [ "采用 Meta 官方 sku_list.py 的标准逻辑架构；不将并行 checkpoint 的复制头计入单份载荷。" ], "evidence": "官方架构表 + 公式", "configSha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486", "configUrl": "https://github.com/meta-llama/llama-models/blob/main/models/sku_list.py", "cardUrl": "https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct", "extraSources": [], "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1073741824.0, "arbitrary": 1073741824.0, "checkpoint": 1073741824.0, "checkpointIndependent": 1073741824.0, "globalBytes": 1073741824.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 32768.0, "endpointBytesPerToken": 32768.0, "checkpointBytesPerToken": 32768.0, "checkpointIndependentBytesPerToken": 32768.0, "arbitraryBytesPerToken": 32768.0 }, "defaultRanks": { "endpoint": 34, "checkpoint": 9, "checkpointIndependent": 9, "arbitrary": 4, "growthBytesPerToken": 36, "stateBytes": 1 } }, { "id": "Qwen3.5-4B", "repo": "Qwen/Qwen3.5-4B", "name": "Qwen3.5 4B", "family": "Qwen", "params": "4B", "size": 4, "arch": "GQA + GDN", "tier": "现役代表", "layers": 32, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 16384, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 12582912, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 589824, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-4B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-4B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 8, "kvHeads": 4, "keyDim": 256, "valueDim": 256, "linearLayers": 24, "stateHeads": 32, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "8 × 2 × 4 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1125253120.0, "arbitrary": 1688995889152.0, "checkpoint": 4370464768.0, "checkpointIndependent": 4370464768.0, "globalBytes": 1073741824.0, "windowBytes": 0, "stateBytes": 51511296.0, "recurrentBytes": 50331648.0, "convBytes": 1179648, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 32768.0, "endpointBytesPerToken": 34340.0, "checkpointBytesPerToken": 133376.0, "checkpointIndependentBytesPerToken": 133376.0, "arbitraryBytesPerToken": 51544064.0 }, "defaultRanks": { "endpoint": 35, "checkpoint": 46, "checkpointIndependent": 46, "arbitrary": 75, "growthBytesPerToken": 36, "stateBytes": 75 } }, { "id": "Qwen3.5-9B", "repo": "Qwen/Qwen3.5-9B", "name": "Qwen3.5 9B", "family": "Qwen", "params": "9B", "size": 9, "arch": "GQA + GDN", "tier": "现役代表", "layers": 32, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 16384, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 12582912, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 589824, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "d0883072e01861ed0b2d47be3c16c36a8e81c224c7ffaa310c6558fb3f932b05", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-9B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 8, "kvHeads": 4, "keyDim": 256, "valueDim": 256, "linearLayers": 24, "stateHeads": 32, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "8 × 2 × 4 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1125253120.0, "arbitrary": 1688995889152.0, "checkpoint": 4370464768.0, "checkpointIndependent": 4370464768.0, "globalBytes": 1073741824.0, "windowBytes": 0, "stateBytes": 51511296.0, "recurrentBytes": 50331648.0, "convBytes": 1179648, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 32768.0, "endpointBytesPerToken": 34340.0, "checkpointBytesPerToken": 133376.0, "checkpointIndependentBytesPerToken": 133376.0, "arbitraryBytesPerToken": 51544064.0 }, "defaultRanks": { "endpoint": 35, "checkpoint": 46, "checkpointIndependent": 46, "arbitrary": 75, "growthBytesPerToken": 36, "stateBytes": 75 } }, { "id": "gpt-oss-120b", "repo": "openai/gpt-oss-120b", "name": "gpt-oss-120b", "family": "openai", "params": "约 117B / A5.1B", "size": 117, "arch": "GQA + SWA", "tier": "现役代表", "layers": 36, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 18432, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "SWA K + V", "elements": 18432, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" } ], "notes": [ "MXFP4 是发布权重的精度，不代表 KV 是 4-bit；此处独立选择 KV 精度。" ], "extraSources": [], "configSha256": "933aeb666a3fd851133ddd7686414f369bc564c4185fb5704416550879f10566", "configUrl": "https://huggingface.co/openai/gpt-oss-120b/blob/main/config.json", "cardUrl": "https://huggingface.co/openai/gpt-oss-120b", "evidence": "配置 + 公式", "geometry": { "fullLayers": 18, "windowLayers": 18, "kvHeads": 8, "keyDim": 64, "valueDim": 64, "window": 128 }, "formula": "2 × 8 × 64 × b × (18T + 18min(T,128))", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1212678144.0, "arbitrary": 2415919104.0, "checkpoint": 1509949440.0, "checkpointIndependent": 1509949440.0, "globalBytes": 1207959552.0, "windowBytes": 4718592.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 36864.0, "endpointBytesPerToken": 37008.0, "checkpointBytesPerToken": 46080.0, "checkpointIndependentBytesPerToken": 46080.0, "arbitraryBytesPerToken": 73728.0 }, "defaultRanks": { "endpoint": 40, "checkpoint": 12, "checkpointIndependent": 12, "arbitrary": 19, "growthBytesPerToken": 40, "stateBytes": 1 } }, { "id": "MiMo-V2.6-Pro-RL", "repo": "XiaomiMiMo/MiMo-V2.6-Pro-RL", "name": "MiMo V2.6 Pro · RL", "family": "XiaomiMiMo", "params": "1.02T / A42B", "size": 1020, "arch": "GQA + SWA", "tier": "现役代表", "layers": 70, "maxContext": 1048576, "components": [ { "label": "全局不等维 K + V", "elements": 25600, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "SWA 不等维 K + V", "elements": 153600, "kind": "window", "ratio": 1, "window": 128, "role": "main", "dtype": "selected" } ], "notes": [ "K 为 192 维、V 为 128 维，不能使用 2×192 代替 K+V。" ], "extraSources": [ { "title": "MiMo 混合层实现", "url": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/modeling_mimo_v2.py" } ], "configSha256": "a9dc00e2a0c2c172efd198f42e81e41c95472021b57c65366c06587d44038b62", "configUrl": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/config.json", "cardUrl": "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL", "evidence": "配置 + 公式", "geometry": { "fullLayers": 10, "windowLayers": 60, "kvHeads": 8, "keyDim": 192, "valueDim": 128, "localKvHeads": 8, "localKeyDim": 192, "localValueDim": 128, "window": 128 }, "formula": "8 × (192+128) × b × (10T + 60min(T,128))", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1717043200.0, "arbitrary": 11744051200.0, "checkpoint": 4194304000.0, "checkpointIndependent": 4194304000.0, "globalBytes": 1677721600.0, "windowBytes": 39321600.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 51200.0, "endpointBytesPerToken": 52400.0, "checkpointBytesPerToken": 128000.0, "checkpointIndependentBytesPerToken": 128000.0, "arbitraryBytesPerToken": 358400.0 }, "defaultRanks": { "endpoint": 45, "checkpoint": 40, "checkpointIndependent": 40, "arbitrary": 63, "growthBytesPerToken": 45, "stateBytes": 1 } }, { "id": "Qwen2.5-7B-Instruct", "repo": "Qwen/Qwen2.5-7B-Instruct", "name": "Qwen2.5 7B", "family": "Qwen", "params": "7B", "size": 7, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 28, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 28672, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "配置中的滑窗大于或等于本 checkpoint 的最大上下文；在有效范围内按全长 KV 计算。" ], "extraSources": [], "configSha256": "7463bb0ea78315365e6c6b74de4e73bbcc8359dfb0c5a737584e077d42c0b03c", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct", "evidence": "配置 + 公式", "geometry": { "fullLayers": 28, "kvHeads": 4, "keyDim": 128, "valueDim": 128, "queryHeads": 28 }, "formula": "2 × 28 层 × 4 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1879048192.0, "arbitrary": 1879048192.0, "checkpoint": 1879048192.0, "checkpointIndependent": 1879048192.0, "globalBytes": 1879048192.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 57344.0, "endpointBytesPerToken": 57344.0, "checkpointBytesPerToken": 57344.0, "checkpointIndependentBytesPerToken": 57344.0, "arbitraryBytesPerToken": 57344.0 }, "defaultRanks": { "endpoint": 47, "checkpoint": 16, "checkpointIndependent": 16, "arbitrary": 9, "growthBytesPerToken": 47, "stateBytes": 1 } }, { "id": "Qwen3.5-27B", "repo": "Qwen/Qwen3.5-27B", "name": "Qwen3.5 27B", "family": "Qwen", "params": "27B", "size": 27, "arch": "GQA + GDN", "tier": "现役代表", "layers": 64, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 32768, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 37748736, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 1474560, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "f8d190c5b89c1521220f935d2567a587d6e291ed69066a45a106560b05a2174c", "configUrl": "https://huggingface.co/Qwen/Qwen3.5-27B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.5-27B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 16, "kvHeads": 4, "keyDim": 256, "valueDim": 256, "linearLayers": 48, "stateHeads": 48, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "16 × 2 × 4 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2301427712.0, "arbitrary": 5046586572800.0, "checkpoint": 11999903744.0, "checkpointIndependent": 11999903744.0, "globalBytes": 2147483648.0, "windowBytes": 0, "stateBytes": 153944064.0, "recurrentBytes": 150994944.0, "convBytes": 2949120, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 65536.0, "endpointBytesPerToken": 70234.0, "checkpointBytesPerToken": 366208.0, "checkpointIndependentBytesPerToken": 366208.0, "arbitraryBytesPerToken": 154009600.0 }, "defaultRanks": { "endpoint": 51, "checkpoint": 90, "checkpointIndependent": 87, "arbitrary": 93, "growthBytesPerToken": 49, "stateBytes": 93 } }, { "id": "DeepSeek-V3", "repo": "deepseek-ai/DeepSeek-V3", "name": "DeepSeek V3", "family": "deepseek-ai", "params": "671B / A37B", "size": 671, "arch": "MLA", "tier": "现役代表", "layers": 61, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 35136, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。" ], "extraSources": [], "configSha256": "cbf0b95dc614de208a109bb5fd4e7eed11385e9c68411d2c17db5319443035d9", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V3", "evidence": "配置 + 公式", "geometry": { "fullLayers": 61, "latentDim": 512, "positionalDim": 64 }, "formula": "61 × (512+64) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2302672896.0, "arbitrary": 2302672896.0, "checkpoint": 2302672896.0, "checkpointIndependent": 2302672896.0, "globalBytes": 2302672896.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 70272.0, "endpointBytesPerToken": 70272.0, "checkpointBytesPerToken": 70272.0, "checkpointIndependentBytesPerToken": 70272.0, "arbitraryBytesPerToken": 70272.0 }, "defaultRanks": { "endpoint": 53, "checkpoint": 22, "checkpointIndependent": 22, "arbitrary": 15, "growthBytesPerToken": 54, "stateBytes": 1 } }, { "id": "DeepSeek-R1", "repo": "deepseek-ai/DeepSeek-R1", "name": "DeepSeek R1", "family": "deepseek-ai", "params": "671B / A37B", "size": 671, "arch": "MLA", "tier": "现役代表", "layers": 61, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 35136, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。" ], "extraSources": [], "configSha256": "79ddea672a62e95d3f0be27be434375538e1988975971cffcf87c96c1de84a65", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-R1/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-R1", "evidence": "配置 + 公式", "geometry": { "fullLayers": 61, "latentDim": 512, "positionalDim": 64 }, "formula": "61 × (512+64) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2302672896.0, "arbitrary": 2302672896.0, "checkpoint": 2302672896.0, "checkpointIndependent": 2302672896.0, "globalBytes": 2302672896.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 70272.0, "endpointBytesPerToken": 70272.0, "checkpointBytesPerToken": 70272.0, "checkpointIndependentBytesPerToken": 70272.0, "arbitraryBytesPerToken": 70272.0 }, "defaultRanks": { "endpoint": 53, "checkpoint": 22, "checkpointIndependent": 22, "arbitrary": 15, "growthBytesPerToken": 54, "stateBytes": 1 } }, { "id": "Kimi-K2.5", "repo": "moonshotai/Kimi-K2.5", "name": "Kimi K2.5", "family": "moonshotai", "params": "1T / A32B", "size": 1000, "arch": "MLA", "tier": "现役代表", "layers": 61, "maxContext": 262144, "components": [ { "label": "MLA latent + 位置分支", "elements": 35136, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。" ], "extraSources": [], "configSha256": "acd5bb01a16f64b309599cd6ed196be056f613c99d6bc9300692b82cd10882f6", "configUrl": "https://huggingface.co/moonshotai/Kimi-K2.5/blob/main/config.json", "cardUrl": "https://huggingface.co/moonshotai/Kimi-K2.5", "evidence": "配置 + 公式", "geometry": { "fullLayers": 61, "latentDim": 512, "positionalDim": 64 }, "formula": "61 × (512+64) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2302672896.0, "arbitrary": 2302672896.0, "checkpoint": 2302672896.0, "checkpointIndependent": 2302672896.0, "globalBytes": 2302672896.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 70272.0, "endpointBytesPerToken": 70272.0, "checkpointBytesPerToken": 70272.0, "checkpointIndependentBytesPerToken": 70272.0, "arbitraryBytesPerToken": 70272.0 }, "defaultRanks": { "endpoint": 53, "checkpoint": 22, "checkpointIndependent": 22, "arbitrary": 15, "growthBytesPerToken": 54, "stateBytes": 1 } }, { "id": "Kimi-K2.6", "repo": "moonshotai/Kimi-K2.6", "name": "Kimi K2.6", "family": "moonshotai", "params": "1T / A32B", "size": 1000, "arch": "MLA", "tier": "现役代表", "layers": 61, "maxContext": 262144, "components": [ { "label": "MLA latent + 位置分支", "elements": 35136, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。" ], "extraSources": [], "configSha256": "85825ca6e18cbe539eb83ee09eedfb3f4222265929f06e9f535a6d9364f55899", "configUrl": "https://huggingface.co/moonshotai/Kimi-K2.6/blob/main/config.json", "cardUrl": "https://huggingface.co/moonshotai/Kimi-K2.6", "evidence": "配置 + 公式", "geometry": { "fullLayers": 61, "latentDim": 512, "positionalDim": 64 }, "formula": "61 × (512+64) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2302672896.0, "arbitrary": 2302672896.0, "checkpoint": 2302672896.0, "checkpointIndependent": 2302672896.0, "globalBytes": 2302672896.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 70272.0, "endpointBytesPerToken": 70272.0, "checkpointBytesPerToken": 70272.0, "checkpointIndependentBytesPerToken": 70272.0, "arbitraryBytesPerToken": 70272.0 }, "defaultRanks": { "endpoint": 53, "checkpoint": 22, "checkpointIndependent": 22, "arbitrary": 15, "growthBytesPerToken": 54, "stateBytes": 1 } }, { "id": "gemma-4-31B", "repo": "google/gemma-4-31B", "name": "Gemma 4 31B", "family": "google", "params": "30.7B", "size": 30.7, "arch": "GQA + SWA", "tier": "现役代表", "layers": 60, "maxContext": 262144, "components": [ { "label": "全局 K + V（独立缓存层）", "elements": 40960, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "滑窗 K + V（独立缓存层）", "elements": 409600, "kind": "window", "ratio": 1, "window": 1024, "role": "main", "dtype": "selected" } ], "notes": [ "attention_k_eq_v 是投影权重共享；K 经过 RoPE、V 不经过 RoPE，缓存仍分别保存 K 与 V，不能据此再减半。", "SWA 按执行窗口回收后的驻留量计算；保留所有历史滑窗 KV 的后端会占用更多。" ], "extraSources": [ { "title": "Gemma 4 共享层映射", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/configuration_gemma4.py" }, { "title": "Gemma 4 K/V 缓存实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/gemma4/modeling_gemma4.py" } ], "configSha256": "6a81841cad2b6ba06841e23c6afb5f0a27827bc12c64328ffa6338831a21267e", "configUrl": "https://huggingface.co/google/gemma-4-31B/blob/main/config.json", "cardUrl": "https://huggingface.co/google/gemma-4-31B", "evidence": "配置 + 共享语义核验", "geometry": { "fullLayers": 10, "windowLayers": 50, "sharedLayers": 0, "kvHeads": 4, "keyDim": 512, "valueDim": 512, "localKvHeads": 16, "localHeadDim": 256, "window": 1024 }, "formula": "10 × 2 × 4 × 512 × b × T + 50 × 2 × 16 × 256 × b × min(T,1024)", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3523215360.0, "arbitrary": 29527900160.0, "checkpoint": 29527900160.0, "checkpointIndependent": 55952015360.0, "globalBytes": 2684354560.0, "windowBytes": 838860800.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 81920.0, "endpointBytesPerToken": 107520.0, "checkpointBytesPerToken": 901120.0, "checkpointIndependentBytesPerToken": 1707520.0, "arbitraryBytesPerToken": 901120.0 }, "defaultRanks": { "endpoint": 64, "checkpoint": 98, "checkpointIndependent": 99, "arbitrary": 69, "growthBytesPerToken": 59, "stateBytes": 1 } }, { "id": "DeepSeek-V3.2", "repo": "deepseek-ai/DeepSeek-V3.2", "name": "DeepSeek V3.2", "family": "deepseek-ai", "params": "671B / A37B", "size": 671, "arch": "MLA + DSA", "tier": "现役代表", "layers": 61, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 35136, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "DSA index K", "elements": 7808, "kind": "linear", "ratio": 1, "window": null, "role": "index", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。", "主表把索引也标准化为所选精度；官方参考实现的原生 index K 是 FP8+FP32 scale（每层每 token 128+4 B），并非主表的 BF16 索引。" ], "extraSources": [ { "title": "DeepSeek V3.2 索引缓存", "url": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/inference/model.py" } ], "configSha256": "c7fa8b191e9936d8e6a57d864baab82b792fae16a116416cdd3a75ba76bc5af1", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V3.2", "evidence": "配置 + 索引实现核验", "geometry": { "fullLayers": 61, "latentDim": 512, "positionalDim": 64, "indexLayers": 61, "indexDim": 128 }, "formula": "61 × (512+64) × b × T + 61 × 128 × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2814377984.0, "arbitrary": 2814377984.0, "checkpoint": 2814377984.0, "checkpointIndependent": 2814377984.0, "globalBytes": 2814377984.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 85888.0, "endpointBytesPerToken": 85888.0, "checkpointBytesPerToken": 85888.0, "checkpointIndependentBytesPerToken": 85888.0, "arbitraryBytesPerToken": 85888.0 }, "defaultRanks": { "endpoint": 59, "checkpoint": 29, "checkpointIndependent": 29, "arbitrary": 22, "growthBytesPerToken": 61, "stateBytes": 1 } }, { "id": "Qwen3.8-2.4T-A95B", "repo": "Qwen/Qwen3.8-2.4T-A95B", "name": "Qwen3.8 2.4T-A95B", "family": "Qwen", "params": "2.4T / A95B", "size": 2400, "arch": "GQA + GDN", "tier": "现役代表", "layers": 92, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 47104, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "GDN 递归状态", "elements": 144703488, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "recurrent", "stateClass": "recurrent" }, { "label": "GDN BF16 短卷积历史", "elements": 4239360, "kind": "fixed", "ratio": 1, "window": null, "role": "state", "dtype": "bf16", "stateClass": "convolution" } ], "notes": [ "递归状态默认 FP32；卷积历史按 BF16、K−1 槽规范化。K 槽参考实现、TP 复制及工作副本需另加。" ], "extraSources": [ { "title": "SGLang 状态形状（K−1 卷积槽）", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Transformers GDN 递归实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_5/modeling_qwen3_5.py" } ], "configSha256": "4e3819548967e319ab435d044a3a331dbe3b078590ce822e9d74b79430533987", "configUrl": "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 23, "kvHeads": 4, "keyDim": 256, "valueDim": 256, "linearLayers": 69, "stateHeads": 128, "stateKeyHeads": 16, "stateKeyDim": 128, "stateValueDim": 128, "convHistory": 3 }, "formula": "23 × 2 × 4 × 256 × b × T + S_GDN", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3674300416.0, "arbitrary": 19247493283840.0, "checkpoint": 40673738752.0, "checkpointIndependent": 40673738752.0, "globalBytes": 3087007744.0, "windowBytes": 0, "stateBytes": 587292672.0, "recurrentBytes": 578813952.0, "convBytes": 8478720, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 94208.0, "endpointBytesPerToken": 112130.75, "checkpointBytesPerToken": 1241264.0, "checkpointIndependentBytesPerToken": 1241264.0, "arbitraryBytesPerToken": 587386880.0 }, "defaultRanks": { "endpoint": 66, "checkpoint": 100, "checkpointIndependent": 97, "arbitrary": 100, "growthBytesPerToken": 62, "stateBytes": 100 } }, { "id": "GLM-5.3", "repo": "zai-org/GLM-5.3", "name": "GLM-5.3", "family": "zai-org", "params": "约 753B（发布权重）", "size": 753, "arch": "MLA + DSA", "tier": "现役代表", "layers": 78, "maxContext": 1048576, "components": [ { "label": "MLA latent + 位置分支", "elements": 44928, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "DSA index K", "elements": 2688, "kind": "linear", "ratio": 1, "window": null, "role": "index", "dtype": "selected" } ], "notes": [ "采用吸收投影后的 MLA 压缩缓存；不乘注意力头数，也不额外乘 K/V 的 2。展开 K/V 的参考路径会大得多。", "21 个 full indexer 保存索引 K，57 层复用 top-k 选择；主 MLA KV 仍有 78 层。" ], "extraSources": [ { "title": "GLM DSA 索引跨层复用", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py" } ], "configSha256": "3ac72612095574542f7fff847ada8e59d9199dd8af44bdf625d7e02615572e69", "configUrl": "https://huggingface.co/zai-org/GLM-5.3/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-5.3", "evidence": "配置 + 索引实现核验", "geometry": { "fullLayers": 78, "latentDim": 512, "positionalDim": 64, "indexLayers": 21, "indexDim": 128 }, "formula": "78 × (512+64) × b × T + 21 × 128 × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3120562176.0, "arbitrary": 3120562176.0, "checkpoint": 3120562176.0, "checkpointIndependent": 3120562176.0, "globalBytes": 3120562176.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 95232.0, "endpointBytesPerToken": 95232.0, "checkpointBytesPerToken": 95232.0, "checkpointIndependentBytesPerToken": 95232.0, "arbitraryBytesPerToken": 95232.0 }, "defaultRanks": { "endpoint": 61, "checkpoint": 31, "checkpointIndependent": 31, "arbitrary": 24, "growthBytesPerToken": 63, "stateBytes": 1 } }, { "id": "Qwen3-30B-A3B", "repo": "Qwen/Qwen3-30B-A3B", "name": "Qwen3 30B-A3B", "family": "Qwen", "params": "30B / A3B", "size": 30, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 48, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 49152, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "2850ddb3bf7aecad20b611e2d44f3077fc8193f4827c93beddd4c02ad63c2297", "configUrl": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-30B-A3B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 48, "kvHeads": 4, "keyDim": 128, "valueDim": 128, "queryHeads": 32 }, "formula": "2 × 48 层 × 4 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3221225472.0, "arbitrary": 3221225472.0, "checkpoint": 3221225472.0, "checkpointIndependent": 3221225472.0, "globalBytes": 3221225472.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 98304.0, "endpointBytesPerToken": 98304.0, "checkpointBytesPerToken": 98304.0, "checkpointIndependentBytesPerToken": 98304.0, "arbitraryBytesPerToken": 98304.0 }, "defaultRanks": { "endpoint": 62, "checkpoint": 32, "checkpointIndependent": 32, "arbitrary": 25, "growthBytesPerToken": 64, "stateBytes": 1 } }, { "id": "Qwen3-1.7B", "repo": "Qwen/Qwen3-1.7B", "name": "Qwen3 1.7B", "family": "Qwen", "params": "1.7B", "size": 1.7, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 28, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 57344, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "1ddb5b89ebc90dcb417a45c213d818577e65976454d29385c8f6140771d95197", "configUrl": "https://huggingface.co/Qwen/Qwen3-1.7B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-1.7B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 28, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 16 }, "formula": "2 × 28 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3758096384.0, "arbitrary": 3758096384.0, "checkpoint": 3758096384.0, "checkpointIndependent": 3758096384.0, "globalBytes": 3758096384.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 114688.0, "endpointBytesPerToken": 114688.0, "checkpointBytesPerToken": 114688.0, "checkpointIndependentBytesPerToken": 114688.0, "arbitraryBytesPerToken": 114688.0 }, "defaultRanks": { "endpoint": 67, "checkpoint": 36, "checkpointIndependent": 36, "arbitrary": 28, "growthBytesPerToken": 67, "stateBytes": 1 } }, { "id": "Llama-3.2-3B-Instruct", "repo": "meta-llama/Llama-3.2-3B-Instruct", "name": "Llama 3.2 3B", "family": "meta-llama", "params": "3B", "size": 3, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 28, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 57344, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "geometry": { "fullLayers": 28, "kvHeads": 8, "keyDim": 128, "valueDim": 128 }, "formula": "2 × 28 层 × 8 KV heads × 128 dim × b", "notes": [ "采用 Meta 官方 sku_list.py 的标准逻辑架构；不将并行 checkpoint 的复制头计入单份载荷。" ], "evidence": "官方架构表 + 公式", "configSha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486", "configUrl": "https://github.com/meta-llama/llama-models/blob/main/models/sku_list.py", "cardUrl": "https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct", "extraSources": [], "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3758096384.0, "arbitrary": 3758096384.0, "checkpoint": 3758096384.0, "checkpointIndependent": 3758096384.0, "globalBytes": 3758096384.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 114688.0, "endpointBytesPerToken": 114688.0, "checkpointBytesPerToken": 114688.0, "checkpointIndependentBytesPerToken": 114688.0, "arbitraryBytesPerToken": 114688.0 }, "defaultRanks": { "endpoint": 67, "checkpoint": 36, "checkpointIndependent": 36, "arbitrary": 28, "growthBytesPerToken": 67, "stateBytes": 1 } }, { "id": "Phi-4-mini-instruct", "repo": "microsoft/Phi-4-mini-instruct", "name": "Phi-4 mini", "family": "microsoft", "params": "3.8B", "size": 3.8, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 32, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "配置中的滑窗大于或等于本 checkpoint 的最大上下文；在有效范围内按全长 KV 计算。" ], "extraSources": [], "configSha256": "ac65d86061d3d0d704ee2511fd0eb8713ef19eb6eedba17c3080a4165d5b933b", "configUrl": "https://huggingface.co/microsoft/Phi-4-mini-instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/microsoft/Phi-4-mini-instruct", "evidence": "配置 + 公式", "geometry": { "fullLayers": 32, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 24 }, "formula": "2 × 32 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": 69, "checkpoint": 41, "checkpointIndependent": 41, "arbitrary": 30, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "Llama-3.1-8B-Instruct", "repo": "meta-llama/Llama-3.1-8B-Instruct", "name": "Llama 3.1 8B", "family": "meta-llama", "params": "8B", "size": 8, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 32, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "geometry": { "fullLayers": 32, "kvHeads": 8, "keyDim": 128, "valueDim": 128 }, "formula": "2 × 32 层 × 8 KV heads × 128 dim × b", "notes": [ "采用 Meta 官方 sku_list.py 的标准逻辑架构；不将并行 checkpoint 的复制头计入单份载荷。" ], "evidence": "官方架构表 + 公式", "configSha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486", "configUrl": "https://github.com/meta-llama/llama-models/blob/main/models/sku_list.py", "cardUrl": "https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct", "extraSources": [], "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": 69, "checkpoint": 41, "checkpointIndependent": 41, "arbitrary": 30, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "MiniMax-M3.1-Flash", "repo": "MiniMaxAI/MiniMax-M3.1-Flash", "name": "MiniMax M3.1 Flash", "family": "MiniMaxAI", "params": "约 428B / A23B", "size": 428, "arch": "GQA + 稀疏索引", "tier": "现役代表", "layers": 60, "maxContext": 1048576, "components": [ { "label": "主 K + V", "elements": 61440, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" }, { "label": "单头 index K", "elements": 7680, "kind": "linear", "ratio": 1, "window": null, "role": "index", "dtype": "selected" } ], "notes": [ "稀疏 top-k 减少注意力读取，不自动减少已保存的全历史主 K/V；indexer 的 4 个 query heads 共用 1 个 key head，且本配置禁用 index V。" ], "extraSources": [ { "title": "SGLang MiniMax M3 索引缓存实现", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/minimax_m3.py" } ], "configSha256": "df6a3668e2de5b8dc0e883e3767495ad17af7e4d69fabe9683088be015fda9ab", "configUrl": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash", "evidence": "配置 + 索引实现核验", "geometry": { "fullLayers": 60, "kvHeads": 4, "keyDim": 128, "valueDim": 128, "indexLayers": 60, "indexDim": 128 }, "formula": "60 × (2 × 4 × 128 + 128) × b × T", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4529848320.0, "arbitrary": 4529848320.0, "checkpoint": 4529848320.0, "checkpointIndependent": 4529848320.0, "globalBytes": 4529848320.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 138240.0, "endpointBytesPerToken": 138240.0, "checkpointBytesPerToken": 138240.0, "checkpointIndependentBytesPerToken": 138240.0, "arbitraryBytesPerToken": 138240.0 }, "defaultRanks": { "endpoint": 74, "checkpoint": 48, "checkpointIndependent": 48, "arbitrary": 35, "growthBytesPerToken": 75, "stateBytes": 1 } }, { "id": "Qwen3-4B", "repo": "Qwen/Qwen3-4B", "name": "Qwen3 4B", "family": "Qwen", "params": "4B", "size": 4, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 36, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 73728, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "8ba006f74fecfaaeb392872a60f4a480e7ec9860153d2e1b769ec81f9a147f8a", "configUrl": "https://huggingface.co/Qwen/Qwen3-4B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-4B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 36, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 32 }, "formula": "2 × 36 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4831838208.0, "arbitrary": 4831838208.0, "checkpoint": 4831838208.0, "checkpointIndependent": 4831838208.0, "globalBytes": 4831838208.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 147456.0, "endpointBytesPerToken": 147456.0, "checkpointBytesPerToken": 147456.0, "checkpointIndependentBytesPerToken": 147456.0, "arbitraryBytesPerToken": 147456.0 }, "defaultRanks": { "endpoint": 76, "checkpoint": 52, "checkpointIndependent": 52, "arbitrary": 37, "growthBytesPerToken": 77, "stateBytes": 1 } }, { "id": "Qwen3-8B", "repo": "Qwen/Qwen3-8B", "name": "Qwen3 8B", "family": "Qwen", "params": "8B", "size": 8, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 36, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 73728, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "f7c4eadfbbf522470667b797a3c89be2524832d2d599797248dc304fff447c30", "configUrl": "https://huggingface.co/Qwen/Qwen3-8B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-8B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 36, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 32 }, "formula": "2 × 36 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4831838208.0, "arbitrary": 4831838208.0, "checkpoint": 4831838208.0, "checkpointIndependent": 4831838208.0, "globalBytes": 4831838208.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 147456.0, "endpointBytesPerToken": 147456.0, "checkpointBytesPerToken": 147456.0, "checkpointIndependentBytesPerToken": 147456.0, "arbitraryBytesPerToken": 147456.0 }, "defaultRanks": { "endpoint": 76, "checkpoint": 52, "checkpointIndependent": 52, "arbitrary": 37, "growthBytesPerToken": 77, "stateBytes": 1 } }, { "id": "Qwen3-14B", "repo": "Qwen/Qwen3-14B", "name": "Qwen3 14B", "family": "Qwen", "params": "14B", "size": 14, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 40, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 81920, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "e73c3664ca09b10a673fef0c22e8a6b456201d49bd4713c9691f775720e8857a", "configUrl": "https://huggingface.co/Qwen/Qwen3-14B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-14B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 40, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 40 }, "formula": "2 × 40 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5368709120.0, "arbitrary": 5368709120.0, "checkpoint": 5368709120.0, "checkpointIndependent": 5368709120.0, "globalBytes": 5368709120.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 163840.0, "endpointBytesPerToken": 163840.0, "checkpointBytesPerToken": 163840.0, "checkpointIndependentBytesPerToken": 163840.0, "arbitraryBytesPerToken": 163840.0 }, "defaultRanks": { "endpoint": 79, "checkpoint": 55, "checkpointIndependent": 55, "arbitrary": 40, "growthBytesPerToken": 80, "stateBytes": 1 } }, { "id": "Mistral-Small-3.2-24B-Instruct-2506", "repo": "mistralai/Mistral-Small-3.2-24B-Instruct-2506", "name": "Mistral Small 3.2", "family": "mistralai", "params": "24B", "size": 24, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 81920, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "01ab910a5dda7995709cc355d094eabb8094b78d49240cd167188606c3ff5edb", "configUrl": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Mistral-Small-3.2-24B-Instruct-2506", "evidence": "配置 + 公式", "geometry": { "fullLayers": 40, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 32 }, "formula": "2 × 40 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5368709120.0, "arbitrary": 5368709120.0, "checkpoint": 5368709120.0, "checkpointIndependent": 5368709120.0, "globalBytes": 5368709120.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 163840.0, "endpointBytesPerToken": 163840.0, "checkpointBytesPerToken": 163840.0, "checkpointIndependentBytesPerToken": 163840.0, "arbitraryBytesPerToken": 163840.0 }, "defaultRanks": { "endpoint": 79, "checkpoint": 55, "checkpointIndependent": 55, "arbitrary": 40, "growthBytesPerToken": 80, "stateBytes": 1 } }, { "id": "Qwen3-235B-A22B-Instruct-2507", "repo": "Qwen/Qwen3-235B-A22B-Instruct-2507", "name": "Qwen3 235B-A22B · 2507", "family": "Qwen", "params": "235B / A22B", "size": 235, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 94, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 96256, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "ad59a339665ee1c68cbff06ac72c9e59f5183ebfaf3698c98737e003a870e9e2", "configUrl": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507", "evidence": "配置 + 公式", "geometry": { "fullLayers": 94, "kvHeads": 4, "keyDim": 128, "valueDim": 128, "queryHeads": 64 }, "formula": "2 × 94 层 × 4 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6308233216.0, "arbitrary": 6308233216.0, "checkpoint": 6308233216.0, "checkpointIndependent": 6308233216.0, "globalBytes": 6308233216.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 192512.0, "endpointBytesPerToken": 192512.0, "checkpointBytesPerToken": 192512.0, "checkpointIndependentBytesPerToken": 192512.0, "arbitraryBytesPerToken": 192512.0 }, "defaultRanks": { "endpoint": 85, "checkpoint": 65, "checkpointIndependent": 65, "arbitrary": 46, "growthBytesPerToken": 85, "stateBytes": 1 } }, { "id": "phi-4", "repo": "microsoft/phi-4", "name": "Phi-4", "family": "microsoft", "params": "14B", "size": 14, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 40, "maxContext": 16384, "components": [ { "label": "全长 K + V", "elements": 102400, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "07eedad2c48798b6e3728e4a1b75e0e092019a375ba725a40e40c78d13664045", "configUrl": "https://huggingface.co/microsoft/phi-4/blob/main/config.json", "cardUrl": "https://huggingface.co/microsoft/phi-4", "evidence": "配置 + 公式", "geometry": { "fullLayers": 40, "kvHeads": 10, "keyDim": 128, "valueDim": 128, "queryHeads": 40 }, "formula": "2 × 40 层 × 10 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6710886400.0, "arbitrary": 6710886400.0, "checkpoint": 6710886400.0, "checkpointIndependent": 6710886400.0, "globalBytes": 6710886400.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 204800.0, "endpointBytesPerToken": 204800.0, "checkpointBytesPerToken": 204800.0, "checkpointIndependentBytesPerToken": 204800.0, "arbitraryBytesPerToken": 204800.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "MiniMax-M2.5", "repo": "MiniMaxAI/MiniMax-M2.5", "name": "MiniMax M2.5", "family": "MiniMaxAI", "params": "约 229B（发布权重）", "size": 229, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 62, "maxContext": 196608, "components": [ { "label": "全长 K + V", "elements": 126976, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [ { "title": "官方 Hub 权重参数统计", "url": "https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5" } ], "configSha256": "3372a31ecbdba614b309cfa5b3dfec33ceea3c580cb9612ff35345d92ced8618", "configUrl": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5/blob/main/config.json", "cardUrl": "https://huggingface.co/MiniMaxAI/MiniMax-M2.5", "evidence": "配置 + 公式", "geometry": { "fullLayers": 62, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 48 }, "formula": "2 × 62 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8321499136.0, "arbitrary": 8321499136.0, "checkpoint": 8321499136.0, "checkpointIndependent": 8321499136.0, "globalBytes": 8321499136.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 253952.0, "endpointBytesPerToken": 253952.0, "checkpointBytesPerToken": 253952.0, "checkpointIndependentBytesPerToken": 253952.0, "arbitraryBytesPerToken": 253952.0 }, "defaultRanks": { "endpoint": 90, "checkpoint": 73, "checkpointIndependent": 72, "arbitrary": 54, "growthBytesPerToken": 90, "stateBytes": 1 } }, { "id": "Qwen3-32B", "repo": "Qwen/Qwen3-32B", "name": "Qwen3 32B", "family": "Qwen", "params": "32B", "size": 32, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 64, "maxContext": 40960, "components": [ { "label": "全长 K + V", "elements": 131072, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [], "configSha256": "97e295b63283935788fac5e4f8860862a56d4089538cafc93f0431f2ebe483bb", "configUrl": "https://huggingface.co/Qwen/Qwen3-32B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-32B", "evidence": "配置 + 公式", "geometry": { "fullLayers": 64, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 64 }, "formula": "2 × 64 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8589934592.0, "arbitrary": 8589934592.0, "checkpoint": 8589934592.0, "checkpointIndependent": 8589934592.0, "globalBytes": 8589934592.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 262144.0, "endpointBytesPerToken": 262144.0, "checkpointBytesPerToken": 262144.0, "checkpointIndependentBytesPerToken": 262144.0, "arbitraryBytesPerToken": 262144.0 }, "defaultRanks": { "endpoint": 92, "checkpoint": 75, "checkpointIndependent": 74, "arbitrary": 56, "growthBytesPerToken": 92, "stateBytes": 1 } }, { "id": "Llama-3.3-70B-Instruct", "repo": "meta-llama/Llama-3.3-70B-Instruct", "name": "Llama 3.3 70B", "family": "meta-llama", "params": "70B", "size": 70, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 80, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 163840, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "geometry": { "fullLayers": 80, "kvHeads": 8, "keyDim": 128, "valueDim": 128 }, "formula": "2 × 80 层 × 8 KV heads × 128 dim × b", "notes": [ "采用 Meta 官方 sku_list.py 的标准逻辑架构；不将并行 checkpoint 的复制头计入单份载荷。" ], "evidence": "官方架构表 + 公式", "configSha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486", "configUrl": "https://github.com/meta-llama/llama-models/blob/main/models/sku_list.py", "cardUrl": "https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct", "extraSources": [], "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 10737418240.0, "arbitrary": 10737418240.0, "checkpoint": 10737418240.0, "checkpointIndependent": 10737418240.0, "globalBytes": 10737418240.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 327680.0, "endpointBytesPerToken": 327680.0, "checkpointBytesPerToken": 327680.0, "checkpointIndependentBytesPerToken": 327680.0, "arbitraryBytesPerToken": 327680.0 }, "defaultRanks": { "endpoint": 95, "checkpoint": 84, "checkpointIndependent": 82, "arbitrary": 60, "growthBytesPerToken": 95, "stateBytes": 1 } }, { "id": "Qwen2.5-72B-Instruct", "repo": "Qwen/Qwen2.5-72B-Instruct", "name": "Qwen2.5 72B", "family": "Qwen", "params": "72B", "size": 72, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 80, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 163840, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [ "配置中的滑窗大于或等于本 checkpoint 的最大上下文；在有效范围内按全长 KV 计算。" ], "extraSources": [], "configSha256": "14ca217334fe0fd10148413592d68c99eeb33431ed89c1afa130fee560be2a29", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-72B-Instruct", "evidence": "配置 + 公式", "geometry": { "fullLayers": 80, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 64 }, "formula": "2 × 80 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 10737418240.0, "arbitrary": 10737418240.0, "checkpoint": 10737418240.0, "checkpointIndependent": 10737418240.0, "globalBytes": 10737418240.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 327680.0, "endpointBytesPerToken": 327680.0, "checkpointBytesPerToken": 327680.0, "checkpointIndependentBytesPerToken": 327680.0, "arbitraryBytesPerToken": 327680.0 }, "defaultRanks": { "endpoint": 95, "checkpoint": 84, "checkpointIndependent": 82, "arbitrary": 60, "growthBytesPerToken": 95, "stateBytes": 1 } }, { "id": "GLM-4.7", "repo": "zai-org/GLM-4.7", "name": "GLM-4.7", "family": "zai-org", "params": "约 358B（发布权重）", "size": 358, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 92, "maxContext": 202752, "components": [ { "label": "全长 K + V", "elements": 188416, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "notes": [], "extraSources": [ { "title": "官方 Hub 权重参数统计", "url": "https://huggingface.co/api/models/zai-org/GLM-4.7" } ], "configSha256": "f713924a536ac1965d8af2a3144d3a0ce0acae94f0fbfd7407240cbafcb08c53", "configUrl": "https://huggingface.co/zai-org/GLM-4.7/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-4.7", "evidence": "配置 + 公式", "geometry": { "fullLayers": 92, "kvHeads": 8, "keyDim": 128, "valueDim": 128, "queryHeads": 96 }, "formula": "2 × 92 层 × 8 KV heads × 128 dim × b", "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 12348030976.0, "arbitrary": 12348030976.0, "checkpoint": 12348030976.0, "checkpointIndependent": 12348030976.0, "globalBytes": 12348030976.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 376832.0, "endpointBytesPerToken": 376832.0, "checkpointBytesPerToken": 376832.0, "checkpointIndependentBytesPerToken": 376832.0, "arbitraryBytesPerToken": 376832.0 }, "defaultRanks": { "endpoint": 98, "checkpoint": 92, "checkpointIndependent": 89, "arbitrary": 65, "growthBytesPerToken": 98, "stateBytes": 1 } }, { "id": "Llama-3.1-405B-Instruct", "repo": "meta-llama/Llama-3.1-405B-Instruct", "name": "Llama 3.1 405B", "family": "meta-llama", "params": "405B", "size": 405, "arch": "GQA", "tier": "经典 / 广泛使用", "layers": 126, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 258048, "kind": "linear", "ratio": 1, "window": null, "role": "main", "dtype": "selected" } ], "geometry": { "fullLayers": 126, "kvHeads": 8, "keyDim": 128, "valueDim": 128 }, "formula": "2 × 126 层 × 8 KV heads × 128 dim × b", "notes": [ "采用 Meta 官方 sku_list.py 的标准逻辑架构；不将并行 checkpoint 的复制头计入单份载荷。", "Meta 的 mp16 特定变体列出 16 个 KV 头；这里使用标准 8 个逻辑 KV 头，实际复制需另计。" ], "evidence": "官方架构表 + 公式", "configSha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486", "configUrl": "https://github.com/meta-llama/llama-models/blob/main/models/sku_list.py", "cardUrl": "https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct", "extraSources": [], "partial": false, "new": false, "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 16911433728.0, "arbitrary": 16911433728.0, "checkpoint": 16911433728.0, "checkpointIndependent": 16911433728.0, "globalBytes": 16911433728.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 516096.0, "endpointBytesPerToken": 516096.0, "checkpointBytesPerToken": 516096.0, "checkpointIndependentBytesPerToken": 516096.0, "arbitraryBytesPerToken": 516096.0 }, "defaultRanks": { "endpoint": 100, "checkpoint": 95, "checkpointIndependent": 93, "arbitrary": 67, "growthBytesPerToken": 100, "stateBytes": 1 } }, { "id": "Qwen2.5-3B-Instruct", "repo": "Qwen/Qwen2.5-3B-Instruct", "name": "Qwen2.5 3B", "params": "3B", "size": 3, "arch": "GQA", "family": "Qwen", "layers": 36, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 18432, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 36 }, "notes": [], "extraSources": [], "configSha256": "eed00b17e22553979d090fa492e587e92885e328914c8e0b0b78f0a0d3576b3b", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "36 × 2 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1207959552.0, "arbitrary": 1207959552.0, "checkpoint": 1207959552.0, "checkpointIndependent": 1207959552.0, "globalBytes": 1207959552.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 36864.0, "endpointBytesPerToken": 36864.0, "checkpointBytesPerToken": 36864.0, "checkpointIndependentBytesPerToken": 36864.0, "arbitraryBytesPerToken": 36864.0 }, "defaultRanks": { "endpoint": 38, "checkpoint": 10, "checkpointIndependent": 10, "arbitrary": 5, "growthBytesPerToken": 40, "stateBytes": 1 } }, { "id": "Qwen2.5-14B-Instruct", "repo": "Qwen/Qwen2.5-14B-Instruct", "name": "Qwen2.5 14B", "params": "14B", "size": 14, "arch": "GQA", "family": "Qwen", "layers": 48, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 98304, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 48 }, "notes": [], "extraSources": [], "configSha256": "0f2085dbbe2ee251bd6a6a0797d84a6ce34436044d629aa3cba793b43d311a9e", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-14B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "48 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6442450944.0, "arbitrary": 6442450944.0, "checkpoint": 6442450944.0, "checkpointIndependent": 6442450944.0, "globalBytes": 6442450944.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 196608.0, "endpointBytesPerToken": 196608.0, "checkpointBytesPerToken": 196608.0, "checkpointIndependentBytesPerToken": 196608.0, "arbitraryBytesPerToken": 196608.0 }, "defaultRanks": { "endpoint": 86, "checkpoint": 66, "checkpointIndependent": 66, "arbitrary": 47, "growthBytesPerToken": 86, "stateBytes": 1 } }, { "id": "Qwen2.5-32B-Instruct", "repo": "Qwen/Qwen2.5-32B-Instruct", "name": "Qwen2.5 32B", "params": "32B", "size": 32, "arch": "GQA", "family": "Qwen", "layers": 64, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 131072, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 64 }, "notes": [], "extraSources": [], "configSha256": "9c6772f138ef9e5b3d1c18f2c87e451bbc01f5f1a4eabb36f9bf4f53829b903e", "configUrl": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen2.5-32B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "64 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8589934592.0, "arbitrary": 8589934592.0, "checkpoint": 8589934592.0, "checkpointIndependent": 8589934592.0, "globalBytes": 8589934592.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 262144.0, "endpointBytesPerToken": 262144.0, "checkpointBytesPerToken": 262144.0, "checkpointIndependentBytesPerToken": 262144.0, "arbitraryBytesPerToken": 262144.0 }, "defaultRanks": { "endpoint": 92, "checkpoint": 75, "checkpointIndependent": 74, "arbitrary": 56, "growthBytesPerToken": 92, "stateBytes": 1 } }, { "id": "Qwen3-Coder-480B-A35B-Instruct", "repo": "Qwen/Qwen3-Coder-480B-A35B-Instruct", "name": "Qwen3 Coder 480B-A35B", "params": "480B / A35B", "size": 480, "arch": "GQA", "family": "Qwen", "layers": 62, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 126976, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 62 }, "notes": [], "extraSources": [], "configSha256": "653230e2dd1c41bd1125d046f00259ed98d6294c75c63969f8890c6e139af7de", "configUrl": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "62 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8321499136.0, "arbitrary": 8321499136.0, "checkpoint": 8321499136.0, "checkpointIndependent": 8321499136.0, "globalBytes": 8321499136.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 253952.0, "endpointBytesPerToken": 253952.0, "checkpointBytesPerToken": 253952.0, "checkpointIndependentBytesPerToken": 253952.0, "arbitraryBytesPerToken": 253952.0 }, "defaultRanks": { "endpoint": 90, "checkpoint": 73, "checkpointIndependent": 72, "arbitrary": 54, "growthBytesPerToken": 90, "stateBytes": 1 } }, { "id": "SmolLM2-1.7B-Instruct", "repo": "HuggingFaceTB/SmolLM2-1.7B-Instruct", "name": "SmolLM2 1.7B", "params": "1.7B", "size": 1.7, "arch": "MHA", "family": "HuggingFaceTB", "layers": 24, "maxContext": 8192, "components": [ { "label": "全长 K + V", "elements": 98304, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 32, "keyDim": 64, "valueDim": 64, "fullLayers": 24 }, "notes": [], "extraSources": [], "configSha256": "994f50b16abb4ae00880baefe03c10260b5bd608d2bf586f7056ca05a534feea", "configUrl": "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "24 × 32 × (64+64) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6442450944.0, "arbitrary": 6442450944.0, "checkpoint": 6442450944.0, "checkpointIndependent": 6442450944.0, "globalBytes": 6442450944.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 196608.0, "endpointBytesPerToken": 196608.0, "checkpointBytesPerToken": 196608.0, "checkpointIndependentBytesPerToken": 196608.0, "arbitraryBytesPerToken": 196608.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "SmolLM3-3B", "repo": "HuggingFaceTB/SmolLM3-3B", "name": "SmolLM3 3B", "params": "3B", "size": 3, "arch": "GQA", "family": "HuggingFaceTB", "layers": 36, "maxContext": 65536, "components": [ { "label": "全长 K + V", "elements": 36864, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 4, "keyDim": 128, "valueDim": 128, "fullLayers": 36 }, "notes": [], "extraSources": [], "configSha256": "c72b1031274ff4626e434d0019e88e95a767460135db9ee492eb80652b786af1", "configUrl": "https://huggingface.co/HuggingFaceTB/SmolLM3-3B/blob/main/config.json", "cardUrl": "https://huggingface.co/HuggingFaceTB/SmolLM3-3B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "36 × 4 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2415919104.0, "arbitrary": 2415919104.0, "checkpoint": 2415919104.0, "checkpointIndependent": 2415919104.0, "globalBytes": 2415919104.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 73728.0, "endpointBytesPerToken": 73728.0, "checkpointBytesPerToken": 73728.0, "checkpointIndependentBytesPerToken": 73728.0, "arbitraryBytesPerToken": 73728.0 }, "defaultRanks": { "endpoint": 57, "checkpoint": 27, "checkpointIndependent": 27, "arbitrary": 19, "growthBytesPerToken": 58, "stateBytes": 1 } }, { "id": "OLMo-2-0425-1B", "repo": "allenai/OLMo-2-0425-1B", "name": "OLMo 2 1B", "params": "1B", "size": 1, "arch": "MHA", "family": "allenai", "layers": 16, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 16, "keyDim": 128, "valueDim": 128, "fullLayers": 16 }, "notes": [], "extraSources": [], "configSha256": "3dc581384d61b3366fc34a95d12c20e595014ea62352469c17c3e3a77f32882a", "configUrl": "https://huggingface.co/allenai/OLMo-2-0425-1B/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/OLMo-2-0425-1B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "16 × 16 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "OLMo-2-1124-7B-Instruct", "repo": "allenai/OLMo-2-1124-7B-Instruct", "name": "OLMo 2 7B", "params": "7B", "size": 7, "arch": "MHA", "family": "allenai", "layers": 32, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 262144, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 32, "keyDim": 128, "valueDim": 128, "fullLayers": 32 }, "notes": [], "extraSources": [], "configSha256": "ff8cc8709a229515676797ab6f343a09391041c9a8fbbc78bfec5be4c2e3664e", "configUrl": "https://huggingface.co/allenai/OLMo-2-1124-7B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/OLMo-2-1124-7B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "32 × 32 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 17179869184.0, "arbitrary": 17179869184.0, "checkpoint": 17179869184.0, "checkpointIndependent": 17179869184.0, "globalBytes": 17179869184.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 524288.0, "endpointBytesPerToken": 524288.0, "checkpointBytesPerToken": 524288.0, "checkpointIndependentBytesPerToken": 524288.0, "arbitraryBytesPerToken": 524288.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "OLMo-2-1124-13B-Instruct", "repo": "allenai/OLMo-2-1124-13B-Instruct", "name": "OLMo 2 13B", "params": "13B", "size": 13, "arch": "MHA", "family": "allenai", "layers": 40, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 409600, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 40, "keyDim": 128, "valueDim": 128, "fullLayers": 40 }, "notes": [], "extraSources": [], "configSha256": "10b1de896314ca2dbf5aeda0035113dbeaa34e086f2a29bb4944596a15f4861e", "configUrl": "https://huggingface.co/allenai/OLMo-2-1124-13B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/OLMo-2-1124-13B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "40 × 40 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 26843545600.0, "arbitrary": 26843545600.0, "checkpoint": 26843545600.0, "checkpointIndependent": 26843545600.0, "globalBytes": 26843545600.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 819200.0, "endpointBytesPerToken": 819200.0, "checkpointBytesPerToken": 819200.0, "checkpointIndependentBytesPerToken": 819200.0, "arbitraryBytesPerToken": 819200.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "OLMo-2-0325-32B-Instruct", "repo": "allenai/OLMo-2-0325-32B-Instruct", "name": "OLMo 2 32B", "params": "32B", "size": 32, "arch": "GQA", "family": "allenai", "layers": 64, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 131072, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 64 }, "notes": [], "extraSources": [], "configSha256": "e977b63b4196481a109089de5a4d4479eb18df70297d9c3e18d865cf5b098e51", "configUrl": "https://huggingface.co/allenai/OLMo-2-0325-32B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/OLMo-2-0325-32B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "64 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8589934592.0, "arbitrary": 8589934592.0, "checkpoint": 8589934592.0, "checkpointIndependent": 8589934592.0, "globalBytes": 8589934592.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 262144.0, "endpointBytesPerToken": 262144.0, "checkpointBytesPerToken": 262144.0, "checkpointIndependentBytesPerToken": 262144.0, "arbitraryBytesPerToken": 262144.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "internlm3-8b-instruct", "repo": "internlm/internlm3-8b-instruct", "name": "InternLM3 8B", "params": "8B", "size": 8, "arch": "GQA", "family": "internlm", "layers": 48, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 24576, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 48 }, "notes": [], "extraSources": [], "configSha256": "e1d88b33517b569d39ccb6730548f6eea9adb5dac892d5db74b78f670c0a7471", "configUrl": "https://huggingface.co/internlm/internlm3-8b-instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/internlm/internlm3-8b-instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "48 × 2 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1610612736.0, "arbitrary": 1610612736.0, "checkpoint": 1610612736.0, "checkpointIndependent": 1610612736.0, "globalBytes": 1610612736.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 49152.0, "endpointBytesPerToken": 49152.0, "checkpointBytesPerToken": 49152.0, "checkpointIndependentBytesPerToken": 49152.0, "arbitraryBytesPerToken": 49152.0 }, "defaultRanks": { "endpoint": 42, "checkpoint": 13, "checkpointIndependent": 13, "arbitrary": 6, "growthBytesPerToken": 43, "stateBytes": 1 } }, { "id": "internlm2_5-20b-chat", "repo": "internlm/internlm2_5-20b-chat", "name": "InternLM2.5 20B", "params": "20B", "size": 20, "arch": "GQA", "family": "internlm", "layers": 48, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 98304, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 48 }, "notes": [], "extraSources": [], "configSha256": "0c2a8521c46627f3189d18f804b3709b179b211d43b4527120e0173e0ce4b8cf", "configUrl": "https://huggingface.co/internlm/internlm2_5-20b-chat/blob/main/config.json", "cardUrl": "https://huggingface.co/internlm/internlm2_5-20b-chat", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "48 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6442450944.0, "arbitrary": 6442450944.0, "checkpoint": 6442450944.0, "checkpointIndependent": 6442450944.0, "globalBytes": 6442450944.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 196608.0, "endpointBytesPerToken": 196608.0, "checkpointBytesPerToken": 196608.0, "checkpointIndependentBytesPerToken": 196608.0, "arbitraryBytesPerToken": 196608.0 }, "defaultRanks": { "endpoint": 86, "checkpoint": 66, "checkpointIndependent": 66, "arbitrary": 47, "growthBytesPerToken": 86, "stateBytes": 1 } }, { "id": "GLM-4.5-Air", "repo": "zai-org/GLM-4.5-Air", "name": "GLM-4.5 Air", "params": "106B / A12B", "size": 106, "arch": "GQA", "family": "zai-org", "layers": 46, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 94208, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 46 }, "notes": [], "extraSources": [], "configSha256": "e59f21d7be4b4b917d80bb3a2d1211356d29cf3aeff93372523daf7d0ffb839d", "configUrl": "https://huggingface.co/zai-org/GLM-4.5-Air/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-4.5-Air", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "46 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6174015488.0, "arbitrary": 6174015488.0, "checkpoint": 6174015488.0, "checkpointIndependent": 6174015488.0, "globalBytes": 6174015488.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 188416.0, "endpointBytesPerToken": 188416.0, "checkpointBytesPerToken": 188416.0, "checkpointIndependentBytesPerToken": 188416.0, "arbitraryBytesPerToken": 188416.0 }, "defaultRanks": { "endpoint": 84, "checkpoint": 64, "checkpointIndependent": 64, "arbitrary": 45, "growthBytesPerToken": 84, "stateBytes": 1 } }, { "id": "GLM-4.5", "repo": "zai-org/GLM-4.5", "name": "GLM-4.5", "params": "355B / A32B", "size": 355, "arch": "GQA", "family": "zai-org", "layers": 92, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 188416, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 92 }, "notes": [], "extraSources": [], "configSha256": "a5dbd07371a8a51b0013a63a7a70543a7094af1a9598812b537133e6da78dda7", "configUrl": "https://huggingface.co/zai-org/GLM-4.5/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-4.5", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "92 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 12348030976.0, "arbitrary": 12348030976.0, "checkpoint": 12348030976.0, "checkpointIndependent": 12348030976.0, "globalBytes": 12348030976.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 376832.0, "endpointBytesPerToken": 376832.0, "checkpointBytesPerToken": 376832.0, "checkpointIndependentBytesPerToken": 376832.0, "arbitraryBytesPerToken": 376832.0 }, "defaultRanks": { "endpoint": 98, "checkpoint": 92, "checkpointIndependent": 89, "arbitrary": 65, "growthBytesPerToken": 98, "stateBytes": 1 } }, { "id": "Mistral-7B-Instruct-v0.3", "repo": "mistralai/Mistral-7B-Instruct-v0.3", "name": "Mistral 7B v0.3", "params": "7B", "size": 7, "arch": "GQA", "family": "mistralai", "layers": 32, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 32 }, "notes": [], "extraSources": [], "configSha256": "affafc6478ec0fd07a32f0ca57aa2fc57743f4d17d6730f86a96ac24d1507f99", "configUrl": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "32 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": 69, "checkpoint": 41, "checkpointIndependent": 41, "arbitrary": 30, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "Mixtral-8x7B-Instruct-v0.1", "repo": "mistralai/Mixtral-8x7B-Instruct-v0.1", "name": "Mixtral 8×7B", "params": "46.7B / A12.9B", "size": 46.7, "arch": "GQA", "family": "mistralai", "layers": 32, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 32 }, "notes": [], "extraSources": [], "configSha256": "9d56d04b36d0fd12ff54ae4c5bac769cc176e254e64ff71144614b6318b40793", "configUrl": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "32 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": 69, "checkpoint": 41, "checkpointIndependent": 41, "arbitrary": 30, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "Mixtral-8x22B-Instruct-v0.1", "repo": "mistralai/Mixtral-8x22B-Instruct-v0.1", "name": "Mixtral 8×22B", "params": "141B / A39B", "size": 141, "arch": "GQA", "family": "mistralai", "layers": 56, "maxContext": 65536, "components": [ { "label": "全长 K + V", "elements": 114688, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 56 }, "notes": [], "extraSources": [], "configSha256": "9c4a6138d84029ab666943613e3d5844d2ea8fd6149f44f77188c62e2915e0f5", "configUrl": "https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "56 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 7516192768.0, "arbitrary": 7516192768.0, "checkpoint": 7516192768.0, "checkpointIndependent": 7516192768.0, "globalBytes": 7516192768.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 229376.0, "endpointBytesPerToken": 229376.0, "checkpointBytesPerToken": 229376.0, "checkpointIndependentBytesPerToken": 229376.0, "arbitraryBytesPerToken": 229376.0 }, "defaultRanks": { "endpoint": 89, "checkpoint": 71, "checkpointIndependent": 70, "arbitrary": 53, "growthBytesPerToken": 89, "stateBytes": 1 } }, { "id": "Mistral-Large-Instruct-2411", "repo": "mistralai/Mistral-Large-Instruct-2411", "name": "Mistral Large · 2411", "params": "123B", "size": 123, "arch": "GQA", "family": "mistralai", "layers": 88, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 180224, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 88 }, "notes": [], "extraSources": [], "configSha256": "2c379704c86ffd8273b1aa9beae0ce20c4c36f47a0c86708efb5aa46c0ac7b74", "configUrl": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Mistral-Large-Instruct-2411", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "88 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 11811160064.0, "arbitrary": 11811160064.0, "checkpoint": 11811160064.0, "checkpointIndependent": 11811160064.0, "globalBytes": 11811160064.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 360448.0, "endpointBytesPerToken": 360448.0, "checkpointBytesPerToken": 360448.0, "checkpointIndependentBytesPerToken": 360448.0, "arbitraryBytesPerToken": 360448.0 }, "defaultRanks": { "endpoint": 97, "checkpoint": 89, "checkpointIndependent": 86, "arbitrary": 64, "growthBytesPerToken": 97, "stateBytes": 1 } }, { "id": "Ministral-3-3B-Instruct-2512", "repo": "mistralai/Ministral-3-3B-Instruct-2512", "name": "Ministral 3 3B", "params": "3B", "size": 3, "arch": "GQA", "family": "mistralai", "layers": 26, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 53248, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 26 }, "notes": [], "extraSources": [], "configSha256": "a9c3658d7bb45ab5be77e660d483a9d3f8a3d54bda2c69a1c576bdfefbc16be6", "configUrl": "https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "26 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3489660928.0, "arbitrary": 3489660928.0, "checkpoint": 3489660928.0, "checkpointIndependent": 3489660928.0, "globalBytes": 3489660928.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 106496.0, "endpointBytesPerToken": 106496.0, "checkpointBytesPerToken": 106496.0, "checkpointIndependentBytesPerToken": 106496.0, "arbitraryBytesPerToken": 106496.0 }, "defaultRanks": { "endpoint": 63, "checkpoint": 34, "checkpointIndependent": 34, "arbitrary": 26, "growthBytesPerToken": 65, "stateBytes": 1 } }, { "id": "Ministral-3-8B-Instruct-2512", "repo": "mistralai/Ministral-3-8B-Instruct-2512", "name": "Ministral 3 8B", "params": "8B", "size": 8, "arch": "GQA", "family": "mistralai", "layers": 34, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 69632, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 34 }, "notes": [], "extraSources": [], "configSha256": "196c38570bb4bc8168a59cc0189c9a9b9974be8dcca6b46dfc109dbd250c01aa", "configUrl": "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "34 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4563402752.0, "arbitrary": 4563402752.0, "checkpoint": 4563402752.0, "checkpointIndependent": 4563402752.0, "globalBytes": 4563402752.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 139264.0, "endpointBytesPerToken": 139264.0, "checkpointBytesPerToken": 139264.0, "checkpointIndependentBytesPerToken": 139264.0, "arbitraryBytesPerToken": 139264.0 }, "defaultRanks": { "endpoint": 75, "checkpoint": 49, "checkpointIndependent": 49, "arbitrary": 36, "growthBytesPerToken": 76, "stateBytes": 1 } }, { "id": "Ministral-3-14B-Instruct-2512", "repo": "mistralai/Ministral-3-14B-Instruct-2512", "name": "Ministral 3 14B", "params": "14B", "size": 14, "arch": "GQA", "family": "mistralai", "layers": 40, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 81920, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 40 }, "notes": [], "extraSources": [], "configSha256": "fc51295bdc0b39c015c9c3ca5d649de183eae73a23b8cbab1a0e055be8d5fa04", "configUrl": "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512/blob/main/config.json", "cardUrl": "https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "40 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5368709120.0, "arbitrary": 5368709120.0, "checkpoint": 5368709120.0, "checkpointIndependent": 5368709120.0, "globalBytes": 5368709120.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 163840.0, "endpointBytesPerToken": 163840.0, "checkpointBytesPerToken": 163840.0, "checkpointIndependentBytesPerToken": 163840.0, "arbitraryBytesPerToken": 163840.0 }, "defaultRanks": { "endpoint": 79, "checkpoint": 55, "checkpointIndependent": 55, "arbitrary": 40, "growthBytesPerToken": 80, "stateBytes": 1 } }, { "id": "Hunyuan-A13B-Instruct", "repo": "tencent/Hunyuan-A13B-Instruct", "name": "Hunyuan A13B", "params": "80B / A13B", "size": 80, "arch": "GQA", "family": "tencent", "layers": 32, "maxContext": 32768, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 32 }, "notes": [ "use_cla=false、use_mla=false，因此不按配置中未启用的跨层共享因子除 2。" ], "extraSources": [], "configSha256": "5fb4af8fa99a28c0bfd1ae009b9488acb2c4365956ffab472c38ae2c1a3d378e", "configUrl": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/tencent/Hunyuan-A13B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "32 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 4294967296.0, "arbitrary": 4294967296.0, "checkpoint": 4294967296.0, "checkpointIndependent": 4294967296.0, "globalBytes": 4294967296.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 131072.0, "checkpointBytesPerToken": 131072.0, "checkpointIndependentBytesPerToken": 131072.0, "arbitraryBytesPerToken": 131072.0 }, "defaultRanks": { "endpoint": 69, "checkpoint": 41, "checkpointIndependent": 41, "arbitrary": 30, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "ERNIE-4.5-21B-A3B-PT", "repo": "baidu/ERNIE-4.5-21B-A3B-PT", "name": "ERNIE 4.5 21B-A3B", "params": "21B / A3B", "size": 21, "arch": "GQA", "family": "baidu", "layers": 28, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 28672, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 4, "keyDim": 128, "valueDim": 128, "fullLayers": 28 }, "notes": [], "extraSources": [], "configSha256": "7383c7387390746c1fa0327f8cab3d5c6b078a60bbf0841d4f42277eece5a5f6", "configUrl": "https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT/blob/main/config.json", "cardUrl": "https://huggingface.co/baidu/ERNIE-4.5-21B-A3B-PT", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "28 × 4 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1879048192.0, "arbitrary": 1879048192.0, "checkpoint": 1879048192.0, "checkpointIndependent": 1879048192.0, "globalBytes": 1879048192.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 57344.0, "endpointBytesPerToken": 57344.0, "checkpointBytesPerToken": 57344.0, "checkpointIndependentBytesPerToken": 57344.0, "arbitraryBytesPerToken": 57344.0 }, "defaultRanks": { "endpoint": 47, "checkpoint": 16, "checkpointIndependent": 16, "arbitrary": 9, "growthBytesPerToken": 47, "stateBytes": 1 } }, { "id": "ERNIE-4.5-300B-A47B-PT", "repo": "baidu/ERNIE-4.5-300B-A47B-PT", "name": "ERNIE 4.5 300B-A47B", "params": "300B / A47B", "size": 300, "arch": "GQA", "family": "baidu", "layers": 54, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 110592, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 54 }, "notes": [], "extraSources": [], "configSha256": "607db524c1233d2687e10df8cb049bcc781a5ba232ee3476553450cf9f5bba75", "configUrl": "https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT/blob/main/config.json", "cardUrl": "https://huggingface.co/baidu/ERNIE-4.5-300B-A47B-PT", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "54 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 7247757312.0, "arbitrary": 7247757312.0, "checkpoint": 7247757312.0, "checkpointIndependent": 7247757312.0, "globalBytes": 7247757312.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 221184.0, "endpointBytesPerToken": 221184.0, "checkpointBytesPerToken": 221184.0, "checkpointIndependentBytesPerToken": 221184.0, "arbitraryBytesPerToken": 221184.0 }, "defaultRanks": { "endpoint": 88, "checkpoint": 69, "checkpointIndependent": 69, "arbitrary": 50, "growthBytesPerToken": 88, "stateBytes": 1 } }, { "id": "granite-4.2-3b", "repo": "ibm-granite/granite-4.2-3b", "name": "Granite 4.2 3B", "params": "3B", "size": 3, "arch": "GQA", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 40960, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 64, "valueDim": 64, "fullLayers": 40 }, "notes": [], "extraSources": [], "configSha256": "e4f44b9c6924a283f1d1d36057a89f1a04404bf1561fc2c0b64cf98180e72915", "configUrl": "https://huggingface.co/ibm-granite/granite-4.2-3b/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.2-3b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "40 × 8 × (64+64) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2684354560.0, "arbitrary": 2684354560.0, "checkpoint": 2684354560.0, "checkpointIndependent": 2684354560.0, "globalBytes": 2684354560.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 81920.0, "endpointBytesPerToken": 81920.0, "checkpointBytesPerToken": 81920.0, "checkpointIndependentBytesPerToken": 81920.0, "arbitraryBytesPerToken": 81920.0 }, "defaultRanks": { "endpoint": 58, "checkpoint": 28, "checkpointIndependent": 28, "arbitrary": 21, "growthBytesPerToken": 59, "stateBytes": 1 } }, { "id": "granite-4.2-8b", "repo": "ibm-granite/granite-4.2-8b", "name": "Granite 4.2 8B", "params": "8B", "size": 8, "arch": "GQA", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 81920, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 40 }, "notes": [], "extraSources": [], "configSha256": "85611f4e34633d4e148e6a5f64bc2d3a23ebbc2b014bc87fbda764eecf261cd9", "configUrl": "https://huggingface.co/ibm-granite/granite-4.2-8b/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.2-8b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "40 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5368709120.0, "arbitrary": 5368709120.0, "checkpoint": 5368709120.0, "checkpointIndependent": 5368709120.0, "globalBytes": 5368709120.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 163840.0, "endpointBytesPerToken": 163840.0, "checkpointBytesPerToken": 163840.0, "checkpointIndependentBytesPerToken": 163840.0, "arbitraryBytesPerToken": 163840.0 }, "defaultRanks": { "endpoint": 79, "checkpoint": 55, "checkpointIndependent": 55, "arbitrary": 40, "growthBytesPerToken": 80, "stateBytes": 1 } }, { "id": "granite-4.2-30b", "repo": "ibm-granite/granite-4.2-30b", "name": "Granite 4.2 30B", "params": "30B", "size": 30, "arch": "GQA", "family": "ibm-granite", "layers": 64, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 131072, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 64 }, "notes": [], "extraSources": [], "configSha256": "82c834ac6bf6a2ecbf9801b49c3fb927008df24ea415c35e931b6091cab21277", "configUrl": "https://huggingface.co/ibm-granite/granite-4.2-30b/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.2-30b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "64 × 8 × (128+128) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 8589934592.0, "arbitrary": 8589934592.0, "checkpoint": 8589934592.0, "checkpointIndependent": 8589934592.0, "globalBytes": 8589934592.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 262144.0, "endpointBytesPerToken": 262144.0, "checkpointBytesPerToken": 262144.0, "checkpointIndependentBytesPerToken": 262144.0, "arbitraryBytesPerToken": 262144.0 }, "defaultRanks": { "endpoint": 92, "checkpoint": 75, "checkpointIndependent": 74, "arbitrary": 56, "growthBytesPerToken": 92, "stateBytes": 1 } }, { "id": "Olmo-3-7B-Instruct", "repo": "allenai/Olmo-3-7B-Instruct", "name": "Olmo 3 7B", "params": "7B", "size": 7, "arch": "MHA + SWA", "family": "allenai", "layers": 32, "maxContext": 65536, "components": [ { "label": "全长 K + V", "elements": 65536, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "SWA K + V", "elements": 196608, "kind": "window", "dtype": "selected", "role": "main", "ratio": 1, "window": 4096, "stateClass": null } ], "geometry": { "kvHeads": 32, "keyDim": 128, "valueDim": 128, "fullLayers": 8, "windowLayers": 24 }, "notes": [], "extraSources": [], "configSha256": "6282ae812f9e494bd212e9a3233d68dc4d5f1e2569642423d2b79a3e80373ef8", "configUrl": "https://huggingface.co/allenai/Olmo-3-7B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/Olmo-3-7B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "2 × 32 × 128 × b × (8T + 24min(T,4096))", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5905580032.0, "arbitrary": 17179869184.0, "checkpoint": 17179869184.0, "checkpointIndependent": 101737037824.0, "globalBytes": 4294967296.0, "windowBytes": 1610612736.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 131072.0, "endpointBytesPerToken": 180224.0, "checkpointBytesPerToken": 524288.0, "checkpointIndependentBytesPerToken": 3104768.0, "arbitraryBytesPerToken": 524288.0 }, "defaultRanks": { "endpoint": 83, "checkpoint": 96, "checkpointIndependent": 100, "arbitrary": 68, "growthBytesPerToken": 69, "stateBytes": 1 } }, { "id": "Olmo-3.1-32B-Instruct", "repo": "allenai/Olmo-3.1-32B-Instruct", "name": "Olmo 3.1 32B", "params": "32B", "size": 32, "arch": "GQA + SWA", "family": "allenai", "layers": 64, "maxContext": 65536, "components": [ { "label": "全长 K + V", "elements": 32768, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "SWA K + V", "elements": 98304, "kind": "window", "dtype": "selected", "role": "main", "ratio": 1, "window": 4096, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 16, "windowLayers": 48 }, "notes": [], "extraSources": [], "configSha256": "d47a120cb2669cd46482d5c2da9488903b67759ca167cd85c1127b2e05755f95", "configUrl": "https://huggingface.co/allenai/Olmo-3.1-32B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/allenai/Olmo-3.1-32B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "2 × 8 × 128 × b × (16T + 48min(T,4096))", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2952790016.0, "arbitrary": 8589934592.0, "checkpoint": 8589934592.0, "checkpointIndependent": 50868518912.0, "globalBytes": 2147483648.0, "windowBytes": 805306368.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 65536.0, "endpointBytesPerToken": 90112.0, "checkpointBytesPerToken": 262144.0, "checkpointIndependentBytesPerToken": 1552384.0, "arbitraryBytesPerToken": 262144.0 }, "defaultRanks": { "endpoint": 60, "checkpoint": 75, "checkpointIndependent": 98, "arbitrary": 56, "growthBytesPerToken": 49, "stateBytes": 1 } }, { "id": "Qwen3.6-27B", "repo": "Qwen/Qwen3.6-27B", "name": "Qwen3.6 27B", "params": "27B", "size": 27, "arch": "GQA + GDN", "family": "Qwen", "layers": 64, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 32768, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "GDN 递归矩阵", "elements": 37748736, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "GDN 卷积历史", "elements": 1474560, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 4, "keyDim": 256, "valueDim": 256, "fullLayers": 16, "recurrentLayers": 48, "stateHeads": 48, "stateKeyDim": 128, "stateValueDim": 128, "convChannels": 10240, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" } ], "configSha256": "69db4eb7196bc8190813231b3018ca05d8c2e3abc7b1af19d55c157af44a9d9c", "configUrl": "https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.6-27B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "16 × 2 × 4 × 256 × b × T + S_GDN", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2301427712.0, "arbitrary": 5046586572800.0, "checkpoint": 11999903744.0, "checkpointIndependent": 11999903744.0, "globalBytes": 2147483648.0, "windowBytes": 0, "stateBytes": 153944064.0, "recurrentBytes": 150994944.0, "convBytes": 2949120, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 65536.0, "endpointBytesPerToken": 70234.0, "checkpointBytesPerToken": 366208.0, "checkpointIndependentBytesPerToken": 366208.0, "arbitraryBytesPerToken": 154009600.0 }, "defaultRanks": { "endpoint": 51, "checkpoint": 90, "checkpointIndependent": 87, "arbitrary": 93, "growthBytesPerToken": 49, "stateBytes": 93 } }, { "id": "Qwen3.6-35B-A3B", "repo": "Qwen/Qwen3.6-35B-A3B", "name": "Qwen3.6 35B A3B", "params": "35B / A3B", "size": 35, "arch": "GQA + GDN", "family": "Qwen", "layers": 40, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 10240, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "GDN 递归矩阵", "elements": 15728640, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "GDN 卷积历史", "elements": 737280, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 256, "valueDim": 256, "fullLayers": 10, "recurrentLayers": 30, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 128, "convChannels": 8192, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" } ], "configSha256": "93a4693fa9d8392fbfccd4b3c9873f4bfdcb14fdede978b123d07d19675efe99", "configUrl": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3.6-35B-A3B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "10 × 2 × 2 × 256 × b × T + S_GDN", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 735477760.0, "arbitrary": 2110573772800.0, "checkpoint": 4791992320.0, "checkpointIndependent": 4791992320.0, "globalBytes": 671088640.0, "windowBytes": 0, "stateBytes": 64389120.0, "recurrentBytes": 62914560.0, "convBytes": 1474560, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 20480.0, "endpointBytesPerToken": 22445.0, "checkpointBytesPerToken": 146240.0, "checkpointIndependentBytesPerToken": 146240.0, "arbitraryBytesPerToken": 64409600.0 }, "defaultRanks": { "endpoint": 21, "checkpoint": 50, "checkpointIndependent": 50, "arbitrary": 79, "growthBytesPerToken": 22, "stateBytes": 79 } }, { "id": "Qwen3-Next-80B-A3B-Instruct", "repo": "Qwen/Qwen3-Next-80B-A3B-Instruct", "name": "Qwen3 Next 80B A3B", "params": "80B / A3B", "size": 80, "arch": "GQA + GDN", "family": "Qwen", "layers": 48, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "GDN 递归矩阵", "elements": 18874368, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "GDN 卷积历史", "elements": 884736, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 256, "valueDim": 256, "fullLayers": 12, "recurrentLayers": 36, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 128, "convChannels": 8192, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" } ], "configSha256": "2d483c7cabad7c8704478ed4038fa7e7b2eff840bc00a118eccbe38e2b488303", "configUrl": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "12 × 2 × 2 × 256 × b × T + S_GDN", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 882573312.0, "arbitrary": 2532688527360.0, "checkpoint": 5750390784.0, "checkpointIndependent": 5750390784.0, "globalBytes": 805306368.0, "windowBytes": 0, "stateBytes": 77266944.0, "recurrentBytes": 75497472.0, "convBytes": 1769472, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 24576.0, "endpointBytesPerToken": 26934.0, "checkpointBytesPerToken": 175488.0, "checkpointIndependentBytesPerToken": 175488.0, "arbitraryBytesPerToken": 77291520.0 }, "defaultRanks": { "endpoint": 28, "checkpoint": 60, "checkpointIndependent": 60, "arbitrary": 83, "growthBytesPerToken": 26, "stateBytes": 83 } }, { "id": "Qwen3-Coder-Next", "repo": "Qwen/Qwen3-Coder-Next", "name": "Qwen3 Coder Next", "params": "80B / A3B", "size": 80, "arch": "GQA + GDN", "family": "Qwen", "layers": 48, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "GDN 递归矩阵", "elements": 18874368, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "GDN 卷积历史", "elements": 884736, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 256, "valueDim": 256, "fullLayers": 12, "recurrentLayers": 36, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 128, "convChannels": 8192, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" } ], "configSha256": "a7b8098d3b05777f12bb5677a26bf1240a1bb09def1b06b29e6be86cae2e84f8", "configUrl": "https://huggingface.co/Qwen/Qwen3-Coder-Next/blob/main/config.json", "cardUrl": "https://huggingface.co/Qwen/Qwen3-Coder-Next", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "12 × 2 × 2 × 256 × b × T + S_GDN", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 882573312.0, "arbitrary": 2532688527360.0, "checkpoint": 5750390784.0, "checkpointIndependent": 5750390784.0, "globalBytes": 805306368.0, "windowBytes": 0, "stateBytes": 77266944.0, "recurrentBytes": 75497472.0, "convBytes": 1769472, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 24576.0, "endpointBytesPerToken": 26934.0, "checkpointBytesPerToken": 175488.0, "checkpointIndependentBytesPerToken": 175488.0, "arbitraryBytesPerToken": 77291520.0 }, "defaultRanks": { "endpoint": 28, "checkpoint": 60, "checkpointIndependent": 60, "arbitrary": 83, "growthBytesPerToken": 26, "stateBytes": 83 } }, { "id": "DeepSeek-V2", "repo": "deepseek-ai/DeepSeek-V2", "name": "DeepSeek V2", "params": "236B / A21B", "size": 236, "arch": "MLA", "family": "deepseek-ai", "layers": 60, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 34560, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "fullLayers": 60, "latentDim": 512, "positionalDim": 64 }, "notes": [], "extraSources": [], "configSha256": "36f572d2909bd108d73a814c585fd61afd56266960a13eb67c713e229d3435d1", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "60 × (512+64) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2264924160.0, "arbitrary": 2264924160.0, "checkpoint": 2264924160.0, "checkpointIndependent": 2264924160.0, "globalBytes": 2264924160.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 69120.0, "endpointBytesPerToken": 69120.0, "checkpointBytesPerToken": 69120.0, "checkpointIndependentBytesPerToken": 69120.0, "arbitraryBytesPerToken": 69120.0 }, "defaultRanks": { "endpoint": 49, "checkpoint": 20, "checkpointIndependent": 20, "arbitrary": 13, "growthBytesPerToken": 52, "stateBytes": 1 } }, { "id": "DeepSeek-V2.5", "repo": "deepseek-ai/DeepSeek-V2.5", "name": "DeepSeek V2.5", "params": "236B / A21B", "size": 236, "arch": "MLA", "family": "deepseek-ai", "layers": 60, "maxContext": 163840, "components": [ { "label": "MLA latent + 位置分支", "elements": 34560, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "fullLayers": 60, "latentDim": 512, "positionalDim": 64 }, "notes": [], "extraSources": [], "configSha256": "4cb37d1e07da958c0367cc9697cec37ae0a80ee9a7c922c37232d798fb311e3e", "configUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2.5/blob/main/config.json", "cardUrl": "https://huggingface.co/deepseek-ai/DeepSeek-V2.5", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "60 × (512+64) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 2264924160.0, "arbitrary": 2264924160.0, "checkpoint": 2264924160.0, "checkpointIndependent": 2264924160.0, "globalBytes": 2264924160.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 69120.0, "endpointBytesPerToken": 69120.0, "checkpointBytesPerToken": 69120.0, "checkpointIndependentBytesPerToken": 69120.0, "arbitraryBytesPerToken": 69120.0 }, "defaultRanks": { "endpoint": 49, "checkpoint": 20, "checkpointIndependent": 20, "arbitrary": 13, "growthBytesPerToken": 52, "stateBytes": 1 } }, { "id": "GLM-4.7-Flash", "repo": "zai-org/GLM-4.7-Flash", "name": "GLM-4.7 Flash", "params": "30B / A3B", "size": 30, "arch": "MLA", "family": "zai-org", "layers": 47, "maxContext": 202752, "components": [ { "label": "MLA latent + 位置分支", "elements": 27072, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "fullLayers": 47, "latentDim": 512, "positionalDim": 64 }, "notes": [], "extraSources": [], "configSha256": "dc9b97c7c9bed726a2e6939da4234d5c43abb3edec8812068c9a1af1dbc13acb", "configUrl": "https://huggingface.co/zai-org/GLM-4.7-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-4.7-Flash", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "47 × (512+64) × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1774190592.0, "arbitrary": 1774190592.0, "checkpoint": 1774190592.0, "checkpointIndependent": 1774190592.0, "globalBytes": 1774190592.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 54144.0, "endpointBytesPerToken": 54144.0, "checkpointBytesPerToken": 54144.0, "checkpointIndependentBytesPerToken": 54144.0, "arbitraryBytesPerToken": 54144.0 }, "defaultRanks": { "endpoint": 46, "checkpoint": 15, "checkpointIndependent": 15, "arbitrary": 8, "growthBytesPerToken": 46, "stateBytes": 1 } }, { "id": "GLM-5", "repo": "zai-org/GLM-5", "name": "GLM-5", "params": "744B / A40B", "size": 744, "arch": "MLA + DSA", "family": "zai-org", "layers": 78, "maxContext": 202752, "components": [ { "label": "MLA latent + 位置分支", "elements": 44928, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "DSA index K", "elements": 9984, "kind": "linear", "dtype": "selected", "role": "index", "ratio": 1, "window": null, "stateClass": null } ], "geometry": { "fullLayers": 78, "latentDim": 512, "positionalDim": 64 }, "notes": [], "extraSources": [ { "title": "GLM DSA 实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/glm_moe_dsa/modeling_glm_moe_dsa.py" } ], "configSha256": "308d3ceebe7bd9c2246d20550d5b793098dd03838d1e504f8236063ae152bcae", "configUrl": "https://huggingface.co/zai-org/GLM-5/blob/main/config.json", "cardUrl": "https://huggingface.co/zai-org/GLM-5", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "78 × (512+64) × b × T + 78 × 128 × b × T", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3598712832.0, "arbitrary": 3598712832.0, "checkpoint": 3598712832.0, "checkpointIndependent": 3598712832.0, "globalBytes": 3598712832.0, "windowBytes": 0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 109824.0, "endpointBytesPerToken": 109824.0, "checkpointBytesPerToken": 109824.0, "checkpointIndependentBytesPerToken": 109824.0, "arbitraryBytesPerToken": 109824.0 }, "defaultRanks": { "endpoint": 65, "checkpoint": 35, "checkpointIndependent": 35, "arbitrary": 27, "growthBytesPerToken": 66, "stateBytes": 1 } }, { "id": "Kimi-Linear-48B-A3B-Instruct", "repo": "moonshotai/Kimi-Linear-48B-A3B-Instruct", "name": "Kimi Linear 48B-A3B", "params": "48B / A3B", "size": 48, "arch": "MLA + KDA", "family": "moonshotai", "layers": 27, "maxContext": 1048576, "components": [ { "label": "MLA latent + 64 维分支", "elements": 4032, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "KDA 递归矩阵", "elements": 10485760, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "KDA 卷积历史", "elements": 737280, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "fullLayers": 7, "latentDim": 512, "positionalDim": 64, "recurrentLayers": 20, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 128, "convChannels": 12288, "convHistory": 3 }, "notes": [ "NoPE 没有删除 64 维投影分支；按压缩 MLA cache 计算，不能用参考路径的展开 K/V 体积。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Kimi Linear 官方模型实现", "url": "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct/blob/main/modeling_kimi.py" } ], "configSha256": "a6ac3c2c4b5aa72370f9727f49ffa4432715d20061889acdb37c688be853096e", "configUrl": "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "7 × (512+64) × b × T + S_KDA", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 307658752.0, "arbitrary": 1422972157952.0, "checkpoint": 3042967552.0, "checkpointIndependent": 3042967552.0, "globalBytes": 264241152.0, "windowBytes": 0, "stateBytes": 43417600.0, "recurrentBytes": 41943040.0, "convBytes": 1474560, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 8064.0, "endpointBytesPerToken": 9389.0, "checkpointBytesPerToken": 92864.0, "checkpointIndependentBytesPerToken": 92864.0, "arbitraryBytesPerToken": 43425664.0 }, "defaultRanks": { "endpoint": 7, "checkpoint": 30, "checkpointIndependent": 30, "arbitrary": 73, "growthBytesPerToken": 7, "stateBytes": 73 } }, { "id": "NVIDIA-Nemotron-3-Nano-4B-FP8", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8", "name": "Nemotron 3 Nano 4B", "params": "4B", "size": 4, "arch": "GQA + Mamba2", "family": "nvidia", "layers": 42, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 8192, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 20643840, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 612864, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 4, "recurrentLayers": 21, "stateHeads": 96, "stateKeyDim": 80, "stateValueDim": 128, "convChannels": 9728, "convHistory": 3 }, "notes": [ "实际 pattern：4 attention、21 Mamba2；MLP / MoE 块无跨 token 动态状态，MTP 排除。Mamba 宽度采用 heads×head_dim，不盲用 hidden_size×expand。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Nemotron 官方 cache / mixer", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8/blob/main/modeling_nemotron_h.py" } ], "configSha256": "1a0a0703a73600ff4b1042dde35cb2137c7e4d76dadc5a88f6f436f2323aa3e9", "configUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8/blob/main/config.json", "cardUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "4 × 2 × 8 × 128 × b × T + S_Mamba2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 620672000.0, "arbitrary": 2746530922496.0, "checkpoint": 5900140544.0, "checkpointIndependent": 5900140544.0, "globalBytes": 536870912.0, "windowBytes": 0, "stateBytes": 83801088.0, "recurrentBytes": 82575360.0, "convBytes": 1225728, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 16384.0, "endpointBytesPerToken": 18941.40625, "checkpointBytesPerToken": 180058.0, "checkpointIndependentBytesPerToken": 180058.0, "arbitraryBytesPerToken": 83817472.0 }, "defaultRanks": { "endpoint": 19, "checkpoint": 62, "checkpointIndependent": 62, "arbitrary": 85, "growthBytesPerToken": 17, "stateBytes": 85 } }, { "id": "NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "repo": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "name": "Nemotron 3 Nano 30B-A3B", "params": "30B / A3B", "size": 30, "arch": "GQA + Mamba2", "family": "nvidia", "layers": 52, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 3072, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 12058624, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 423936, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 6, "recurrentLayers": 23, "stateHeads": 64, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 6144, "convHistory": 3 }, "notes": [ "实际 pattern：6 attention、23 Mamba2；MLP / MoE 块无跨 token 动态状态，MTP 排除。Mamba 宽度采用 heads×head_dim，不盲用 hidden_size×expand。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Nemotron 官方 cache / mixer", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/blob/main/modeling_nemotron_h.py" } ], "configSha256": "dd9fa380ac107b0477db5a26108db9febe6378e7bb3966a107944853ec4f76f8", "configUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/blob/main/config.json", "cardUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "6 × 2 × 2 × 128 × b × T + S_Mamba2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 250408960.0, "arbitrary": 1608532361216.0, "checkpoint": 3342598144.0, "checkpointIndependent": 3342598144.0, "globalBytes": 201326592.0, "windowBytes": 0, "stateBytes": 49082368.0, "recurrentBytes": 48234496.0, "convBytes": 847872, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 6144.0, "endpointBytesPerToken": 7641.875, "checkpointBytesPerToken": 102008.0, "checkpointIndependentBytesPerToken": 102008.0, "arbitraryBytesPerToken": 49088512.0 }, "defaultRanks": { "endpoint": 6, "checkpoint": 33, "checkpointIndependent": 33, "arbitrary": 74, "growthBytesPerToken": 4, "stateBytes": 74 } }, { "id": "NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "repo": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "name": "Nemotron 3 Super 120B-A12B", "params": "120B / A12B", "size": 120, "arch": "GQA + Mamba2", "family": "nvidia", "layers": 88, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 4096, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 41943040, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 1228800, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 8, "recurrentLayers": 40, "stateHeads": 128, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 10240, "convHistory": 3 }, "notes": [ "实际 pattern：8 attention、40 Mamba2；MLP / MoE 块无跨 token 动态状态，MTP 排除。Mamba 宽度采用 heads×head_dim，不盲用 hidden_size×expand。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Nemotron 官方 cache / mixer", "url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/blob/main/modeling_nemotron_h.py" } ], "configSha256": "699f34f0fc645d29ebffa5767fb59e6ae6ec98e3a4605485eb9913256d0df7e6", "configUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16/blob/main/config.json", "cardUrl": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "8 × 2 × 2 × 128 × b × T + S_Mamba2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 438665216.0, "arbitrary": 5578357211136.0, "checkpoint": 11163140096.0, "checkpointIndependent": 11163140096.0, "globalBytes": 268435456.0, "windowBytes": 0, "stateBytes": 170229760.0, "recurrentBytes": 167772160.0, "convBytes": 2457600, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 8192.0, "endpointBytesPerToken": 13387.0, "checkpointBytesPerToken": 340672.0, "checkpointIndependentBytesPerToken": 340672.0, "arbitraryBytesPerToken": 170237952.0 }, "defaultRanks": { "endpoint": 15, "checkpoint": 87, "checkpointIndependent": 85, "arbitrary": 96, "growthBytesPerToken": 8, "stateBytes": 96 } }, { "id": "Falcon-H1-1.5B-Instruct", "repo": "tiiuae/Falcon-H1-1.5B-Instruct", "name": "Falcon H1 1.5B", "params": "1.5B", "size": 1.5, "arch": "GQA ∥ Mamba2", "family": "tiiuae", "layers": 24, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 12288, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 18874368, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 258048, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 24, "recurrentLayers": 24, "stateHeads": 48, "stateKeyDim": 64, "stateValueDim": 256, "convChannels": 3584, "convHistory": 3 }, "notes": [ "每个 block 内 attention 与 Mamba2 并行，两类状态都计入；SSM 维度用 mamba_d_ssm，不把 gated-MLP 通道计入递归矩阵。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Falcon-H1 并行 hybrid / 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon_h1/modeling_falcon_h1.py" } ], "configSha256": "2473f5e02854a2833df973277e6f109d94344e7860fbc78d5ff8b31d08dddaeb", "configUrl": "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "24 × 2 × 2 × 128 × b × T + 24 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 881319936.0, "arbitrary": 2491617902592.0, "checkpoint": 5670174720.0, "checkpointIndependent": 5670174720.0, "globalBytes": 805306368.0, "windowBytes": 0, "stateBytes": 76013568.0, "recurrentBytes": 75497472.0, "convBytes": 516096, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 24576.0, "endpointBytesPerToken": 26895.75, "checkpointBytesPerToken": 173040.0, "checkpointIndependentBytesPerToken": 173040.0, "arbitraryBytesPerToken": 76038144.0 }, "defaultRanks": { "endpoint": 27, "checkpoint": 59, "checkpointIndependent": 59, "arbitrary": 81, "growthBytesPerToken": 26, "stateBytes": 81 } }, { "id": "Falcon-H1-3B-Instruct", "repo": "tiiuae/Falcon-H1-3B-Instruct", "name": "Falcon H1 3B", "params": "3B", "size": 3, "arch": "GQA ∥ Mamba2", "family": "tiiuae", "layers": 32, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 16384, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 33554432, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 442368, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 32, "recurrentLayers": 32, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 256, "convChannels": 4608, "convHistory": 3 }, "notes": [ "每个 block 内 attention 与 Mamba2 并行，两类状态都计入；SSM 维度用 mamba_d_ssm，不把 gated-MLP 通道计入递归矩阵。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Falcon-H1 并行 hybrid / 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon_h1/modeling_falcon_h1.py" } ], "configSha256": "f6faecc746e9bd9450747e36e69f802f39551b3eb74fbbf9bef5a7afddb8c12f", "configUrl": "https://huggingface.co/tiiuae/Falcon-H1-3B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/tiiuae/Falcon-H1-3B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "32 × 2 × 2 × 128 × b × T + 32 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1208844288.0, "arbitrary": 4428111282176.0, "checkpoint": 9720299520.0, "checkpointIndependent": 9720299520.0, "globalBytes": 1073741824.0, "windowBytes": 0, "stateBytes": 135102464.0, "recurrentBytes": 134217728.0, "convBytes": 884736, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 32768.0, "endpointBytesPerToken": 36891.0, "checkpointBytesPerToken": 296640.0, "checkpointIndependentBytesPerToken": 296640.0, "arbitraryBytesPerToken": 135135232.0 }, "defaultRanks": { "endpoint": 39, "checkpoint": 79, "checkpointIndependent": 77, "arbitrary": 88, "growthBytesPerToken": 36, "stateBytes": 88 } }, { "id": "Falcon-H1-7B-Instruct", "repo": "tiiuae/Falcon-H1-7B-Instruct", "name": "Falcon H1 7B", "params": "7B", "size": 7, "arch": "GQA ∥ Mamba2", "family": "tiiuae", "layers": 44, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 22528, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 34603008, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 473088, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 2, "keyDim": 128, "valueDim": 128, "fullLayers": 44, "recurrentLayers": 44, "stateHeads": 24, "stateKeyDim": 128, "stateValueDim": 256, "convChannels": 3584, "convHistory": 3 }, "notes": [ "每个 block 内 attention 与 Mamba2 并行，两类状态都计入；SSM 维度用 mamba_d_ssm，不把 gated-MLP 通道计入递归矩阵。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Falcon-H1 并行 hybrid / 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon_h1/modeling_falcon_h1.py" } ], "configSha256": "6a91400ff4652977d914fab8c495716cd6020f017b578dde89cc9744e4199607", "configUrl": "https://huggingface.co/tiiuae/Falcon-H1-7B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/tiiuae/Falcon-H1-7B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "44 × 2 × 2 × 128 × b × T + 44 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1615753216.0, "arbitrary": 4567966154752.0, "checkpoint": 10395320320.0, "checkpointIndependent": 10395320320.0, "globalBytes": 1476395008.0, "windowBytes": 0, "stateBytes": 139358208.0, "recurrentBytes": 138412032.0, "convBytes": 946176, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 45056.0, "endpointBytesPerToken": 49308.875, "checkpointBytesPerToken": 317240.0, "checkpointIndependentBytesPerToken": 317240.0, "arbitraryBytesPerToken": 139403264.0 }, "defaultRanks": { "endpoint": 43, "checkpoint": 82, "checkpointIndependent": 80, "arbitrary": 89, "growthBytesPerToken": 42, "stateBytes": 89 } }, { "id": "Falcon-H1-34B-Instruct", "repo": "tiiuae/Falcon-H1-34B-Instruct", "name": "Falcon H1 34B", "params": "34B", "size": 34, "arch": "GQA ∥ Mamba2", "family": "tiiuae", "layers": 72, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 73728, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 75497472, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 1105920, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 4, "keyDim": 128, "valueDim": 128, "fullLayers": 72, "recurrentLayers": 72, "stateHeads": 32, "stateKeyDim": 128, "stateValueDim": 256, "convChannels": 5120, "convHistory": 3 }, "notes": [ "每个 block 内 attention 与 Mamba2 并行，两类状态都计入；SSM 维度用 mamba_d_ssm，不把 gated-MLP 通道计入递归矩阵。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Falcon-H1 并行 hybrid / 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/falcon_h1/modeling_falcon_h1.py" } ], "configSha256": "156d23cf8c31a19d1951fb530ce703441b8dba7002af7f6185d26a9d191b1a2f", "configUrl": "https://huggingface.co/tiiuae/Falcon-H1-34B-Instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/tiiuae/Falcon-H1-34B-Instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "72 × 2 × 4 × 128 × b × T + 72 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 5136039936.0, "arbitrary": 9972914061312.0, "checkpoint": 24300748800.0, "checkpointIndependent": 24300748800.0, "globalBytes": 4831838208.0, "windowBytes": 0, "stateBytes": 304201728.0, "recurrentBytes": 301989888.0, "convBytes": 2211840, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 147456.0, "endpointBytesPerToken": 156739.5, "checkpointBytesPerToken": 741600.0, "checkpointIndependentBytesPerToken": 741600.0, "arbitraryBytesPerToken": 304349184.0 }, "defaultRanks": { "endpoint": 78, "checkpoint": 97, "checkpointIndependent": 95, "arbitrary": 98, "growthBytesPerToken": 77, "stateBytes": 98 } }, { "id": "granite-4.0-h-1b", "repo": "ibm-granite/granite-4.0-h-1b", "name": "Granite 4.0 H 1b", "params": "1B", "size": 1, "arch": "GQA + Mamba2", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 4096, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 14155776, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 359424, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 4, "keyDim": 128, "valueDim": 128, "fullLayers": 4, "recurrentLayers": 36, "stateHeads": 48, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 3328, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Granite hybrid 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py" } ], "configSha256": "aafbdabf9b0ff18ab18e6d22b867b1ba006305fd666a8d9f579de7c4a98ca0b8", "configUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-1b/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-1b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "4 层 GQA KV + 36 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 325777408.0, "arbitrary": 1879249518592.0, "checkpoint": 3938320384.0, "checkpointIndependent": 3938320384.0, "globalBytes": 268435456.0, "windowBytes": 0, "stateBytes": 57341952.0, "recurrentBytes": 56623104.0, "convBytes": 718848, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 8192.0, "endpointBytesPerToken": 9941.9375, "checkpointBytesPerToken": 120188.0, "checkpointIndependentBytesPerToken": 120188.0, "arbitraryBytesPerToken": 57350144.0 }, "defaultRanks": { "endpoint": 8, "checkpoint": 38, "checkpointIndependent": 38, "arbitrary": 77, "growthBytesPerToken": 8, "stateBytes": 77 } }, { "id": "granite-4.0-h-micro", "repo": "ibm-granite/granite-4.0-h-micro", "name": "Granite 4.0 H micro", "params": "3B", "size": 3, "arch": "GQA + Mamba2", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 4096, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 18874368, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 470016, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 64, "valueDim": 64, "fullLayers": 4, "recurrentLayers": 36, "stateHeads": 64, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 4352, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Granite hybrid 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py" } ], "configSha256": "bd828b00edcc25f93d6fa55210acebcd058054c4d5b5fa6c5146b0a0d18172ee", "configUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-micro/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-micro", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "4 层 GQA KV + 36 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 344872960.0, "arbitrary": 2504972566528.0, "checkpoint": 5160435712.0, "checkpointIndependent": 5160435712.0, "globalBytes": 268435456.0, "windowBytes": 0, "stateBytes": 76437504.0, "recurrentBytes": 75497472.0, "convBytes": 940032, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 8192.0, "endpointBytesPerToken": 10524.6875, "checkpointBytesPerToken": 157484.0, "checkpointIndependentBytesPerToken": 157484.0, "arbitraryBytesPerToken": 76445696.0 }, "defaultRanks": { "endpoint": 10, "checkpoint": 54, "checkpointIndependent": 54, "arbitrary": 82, "growthBytesPerToken": 8, "stateBytes": 82 } }, { "id": "granite-4.0-h-tiny", "repo": "ibm-granite/granite-4.0-h-tiny", "name": "Granite 4.0 H tiny", "params": "7B / A1B", "size": 7, "arch": "GQA + Mamba2", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 4096, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 14155776, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 359424, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 4, "keyDim": 128, "valueDim": 128, "fullLayers": 4, "recurrentLayers": 36, "stateHeads": 48, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 3328, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Granite hybrid 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py" } ], "configSha256": "bda8fd574ace7d968d82397f59ea6b9a702a077bbeab279a65b9dad7386a82c6", "configUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-tiny/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-tiny", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "4 层 GQA KV + 36 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 325777408.0, "arbitrary": 1879249518592.0, "checkpoint": 3938320384.0, "checkpointIndependent": 3938320384.0, "globalBytes": 268435456.0, "windowBytes": 0, "stateBytes": 57341952.0, "recurrentBytes": 56623104.0, "convBytes": 718848, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 8192.0, "endpointBytesPerToken": 9941.9375, "checkpointBytesPerToken": 120188.0, "checkpointIndependentBytesPerToken": 120188.0, "arbitraryBytesPerToken": 57350144.0 }, "defaultRanks": { "endpoint": 8, "checkpoint": 38, "checkpointIndependent": 38, "arbitrary": 77, "growthBytesPerToken": 8, "stateBytes": 77 } }, { "id": "granite-4.0-h-small", "repo": "ibm-granite/granite-4.0-h-small", "name": "Granite 4.0 H small", "params": "32B / A9B", "size": 32, "arch": "GQA + Mamba2", "family": "ibm-granite", "layers": 40, "maxContext": 131072, "components": [ { "label": "全长 K + V", "elements": 8192, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 37748736, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 912384, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 4, "recurrentLayers": 36, "stateHeads": 128, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 8448, "convHistory": 3 }, "notes": [], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Granite hybrid 状态实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/granitemoehybrid/modeling_granitemoehybrid.py" } ], "configSha256": "8616e9f0b30e6fac9696f7c1e1dbd08f1a850ac4af0de6353f7d6009043702ae", "configUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-small/blob/main/config.json", "cardUrl": "https://huggingface.co/ibm-granite/granite-4.0-h-small", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "4 层 GQA KV + 36 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 689690624.0, "arbitrary": 5008133193728.0, "checkpoint": 10317332480.0, "checkpointIndependent": 10317332480.0, "globalBytes": 536870912.0, "windowBytes": 0, "stateBytes": 152819712.0, "recurrentBytes": 150994944.0, "convBytes": 1824768, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 16384.0, "endpointBytesPerToken": 21047.6875, "checkpointBytesPerToken": 314860.0, "checkpointIndependentBytesPerToken": 314860.0, "arbitraryBytesPerToken": 152836096.0 }, "defaultRanks": { "endpoint": 20, "checkpoint": 81, "checkpointIndependent": 79, "arbitrary": 91, "growthBytesPerToken": 17, "stateBytes": 91 } }, { "id": "Zamba2-1.2B-instruct", "repo": "Zyphra/Zamba2-1.2B-instruct", "name": "Zamba2 1.2B", "params": "1.2B", "size": 1.2, "arch": "MHA + Mamba2", "family": "Zyphra", "layers": 38, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 49152, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 19922944, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 496128, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 32, "keyDim": 128, "valueDim": 128, "fullLayers": 6, "recurrentLayers": 38, "stateHeads": 64, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 4352, "convHistory": 3 }, "notes": [ "共享 Transformer 权重不代表共享 KV：每次 hybrid 调用按自己的 layer_idx 写 cache。所有 block（含 hybrid）都有 Mamba 状态。attention_head_dim 为 2×hidden/heads。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Zamba2 layer_idx cache / shared weights", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/zamba2/modeling_zamba2.py" } ], "configSha256": "3cc6e1fd1266a062390f50cd5e13ae36e2f36ada8dbb3443c284581af5eb355a", "configUrl": "https://huggingface.co/Zyphra/Zamba2-1.2B-instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Zyphra/Zamba2-1.2B-instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "6 × 2 × 32 × 128 × b × T + 38 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 3301909504.0, "arbitrary": 2647075586048.0, "checkpoint": 8385003520.0, "checkpointIndependent": 8385003520.0, "globalBytes": 3221225472.0, "windowBytes": 0, "stateBytes": 80684032.0, "recurrentBytes": 79691776.0, "convBytes": 992256, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 98304.0, "endpointBytesPerToken": 100766.28125, "checkpointBytesPerToken": 255890.0, "checkpointIndependentBytesPerToken": 255890.0, "arbitraryBytesPerToken": 80782336.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "Zamba2-2.7B-instruct", "repo": "Zyphra/Zamba2-2.7B-instruct", "name": "Zamba2 2.7B", "params": "2.7B", "size": 2.7, "arch": "MHA + Mamba2", "family": "Zyphra", "layers": 54, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 92160, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 17694720, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 850176, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 32, "keyDim": 160, "valueDim": 160, "fullLayers": 9, "recurrentLayers": 54, "stateHeads": 80, "stateKeyDim": 64, "stateValueDim": 64, "convChannels": 5248, "convHistory": 3 }, "notes": [ "共享 Transformer 权重不代表共享 KV：每次 hybrid 调用按自己的 layer_idx 写 cache。所有 block（含 hybrid）都有 Mamba 状态。attention_head_dim 为 2×hidden/heads。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Zamba2 layer_idx cache / shared weights", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/zamba2/modeling_zamba2.py" } ], "configSha256": "0497ca3001366ed8eaef1030c84f856fd864f6f5e2ba86cef9ee9b211807df02", "configUrl": "https://huggingface.co/Zyphra/Zamba2-2.7B-instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Zyphra/Zamba2-2.7B-instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "9 × 2 × 32 × 160 × b × T + 54 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 6112276992.0, "arbitrary": 2381039271936.0, "checkpoint": 10678468608.0, "checkpointIndependent": 10678468608.0, "globalBytes": 6039797760.0, "windowBytes": 0, "stateBytes": 72479232.0, "recurrentBytes": 70778880.0, "convBytes": 1700352, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 184320.0, "endpointBytesPerToken": 186531.890625, "checkpointBytesPerToken": 325881.0, "checkpointIndependentBytesPerToken": 325881.0, "arbitraryBytesPerToken": 72663552.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "Zamba2-7B-instruct", "repo": "Zyphra/Zamba2-7B-instruct", "name": "Zamba2 7B", "params": "7B", "size": 7, "arch": "MHA + Mamba2", "family": "Zyphra", "layers": 81, "maxContext": 4096, "components": [ { "label": "全长 K + V", "elements": 186368, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "Mamba2 递归矩阵", "elements": 37158912, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 1804032, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 32, "keyDim": 224, "valueDim": 224, "fullLayers": 13, "recurrentLayers": 81, "stateHeads": 112, "stateKeyDim": 64, "stateValueDim": 64, "convChannels": 7424, "convHistory": 3 }, "notes": [ "共享 Transformer 权重不代表共享 KV：每次 hybrid 调用按自己的 layer_idx 写 cache。所有 block（含 hybrid）都有 Mamba 状态。attention_head_dim 为 2×hidden/heads。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Zamba2 layer_idx cache / shared weights", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/zamba2/modeling_zamba2.py" } ], "configSha256": "51d90f27707cd78e64184c5f5295f5b2dcc967c75f4469d522b198a0f245f5e8", "configUrl": "https://huggingface.co/Zyphra/Zamba2-7B-instruct/blob/main/config.json", "cardUrl": "https://huggingface.co/Zyphra/Zamba2-7B-instruct", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "13 × 2 × 32 × 224 × b × T + 81 层 Mamba2 状态", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 12366056960.0, "arbitrary": 5000935768064.0, "checkpoint": 21957410816.0, "checkpointIndependent": 21957410816.0, "globalBytes": 12213813248.0, "windowBytes": 0, "stateBytes": 152243712.0, "recurrentBytes": 148635648.0, "convBytes": 3608064, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 372736.0, "endpointBytesPerToken": 377382.109375, "checkpointBytesPerToken": 670087.0, "checkpointIndependentBytesPerToken": 670087.0, "arbitraryBytesPerToken": 152616448.0 }, "defaultRanks": { "endpoint": null, "checkpoint": null, "checkpointIndependent": null, "arbitrary": null, "growthBytesPerToken": null, "stateBytes": null } }, { "id": "mamba2-1.3b", "repo": "state-spaces/mamba2-1.3b", "name": "Mamba2 1.3b", "params": "1.3B", "size": 1.3, "arch": "纯 Mamba2", "family": "state-spaces", "layers": 48, "maxContext": null, "components": [ { "label": "Mamba2 递归矩阵", "elements": 25165824, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 626688, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "recurrentLayers": 48, "stateHeads": 64, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 4352, "convHistory": 3 }, "notes": [ "官方配置未提供 max_position_embeddings，架构无随 T 增长的 KV；长上下文效果不能由恒定状态大小推断。Mamba2 构造默认 d_state=128、expand=2、headdim=64、ngroups=1、d_conv=4。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Mamba2 官方默认形状与 cache", "url": "https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba2.py" } ], "configSha256": "204f2030f433f61e53fd9c3639662c6cf231cd390071f6ea38b0890b0e3c04f0", "configUrl": "https://huggingface.co/state-spaces/mamba2-1.3b/blob/main/config.json", "cardUrl": "https://huggingface.co/state-spaces/mamba2-1.3b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "48 × 4096 × 128 × b_state + 48 × (4096+256) × 3 × 2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 101916672.0, "arbitrary": 3339605508096.0, "checkpoint": 6522667008.0, "checkpointIndependent": 6522667008.0, "globalBytes": 0, "windowBytes": 0, "stateBytes": 101916672.0, "recurrentBytes": 100663296.0, "convBytes": 1253376, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 0, "endpointBytesPerToken": 3110.25, "checkpointBytesPerToken": 199056.0, "checkpointIndependentBytesPerToken": 199056.0, "arbitraryBytesPerToken": 101916672.0 }, "defaultRanks": { "endpoint": 1, "checkpoint": 68, "checkpointIndependent": 68, "arbitrary": 86, "growthBytesPerToken": 1, "stateBytes": 86 } }, { "id": "mamba2-2.7b", "repo": "state-spaces/mamba2-2.7b", "name": "Mamba2 2.7b", "params": "2.7B", "size": 2.7, "arch": "纯 Mamba2", "family": "state-spaces", "layers": 64, "maxContext": null, "components": [ { "label": "Mamba2 递归矩阵", "elements": 41943040, "kind": "fixed", "dtype": "recurrent", "role": "state", "ratio": 1, "window": null, "stateClass": "recurrent" }, { "label": "Mamba2 卷积历史", "elements": 1032192, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "recurrentLayers": 64, "stateHeads": 80, "stateKeyDim": 64, "stateValueDim": 128, "convChannels": 5376, "convHistory": 3 }, "notes": [ "官方配置未提供 max_position_embeddings，架构无随 T 增长的 KV；长上下文效果不能由恒定状态大小推断。Mamba2 构造默认 d_state=128、expand=2、headdim=64、ngroups=1、d_conv=4。" ], "extraSources": [ { "title": "SGLang 持久状态形状 / K−1 历史", "url": "https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/configs/mamba_utils.py" }, { "title": "Mamba2 官方默认形状与 cache", "url": "https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba2.py" } ], "configSha256": "0955f133a3eeb902ed11e1a65d0be2934e33a8e1871b6a9685414fa8955e6c86", "configUrl": "https://huggingface.co/state-spaces/mamba2-2.7b/blob/main/config.json", "cardUrl": "https://huggingface.co/state-spaces/mamba2-2.7b", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "64 × 5120 × 128 × b_state + 64 × (5120+256) × 3 × 2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 169836544.0, "arbitrary": 5565203873792.0, "checkpoint": 10869538816.0, "checkpointIndependent": 10869538816.0, "globalBytes": 0, "windowBytes": 0, "stateBytes": 169836544.0, "recurrentBytes": 167772160.0, "convBytes": 2064384, "otherStateBytes": 0.0, "checkpointCount": 64, "growthBytesPerToken": 0, "endpointBytesPerToken": 5183.0, "checkpointBytesPerToken": 331712.0, "checkpointIndependentBytesPerToken": 331712.0, "arbitraryBytesPerToken": 169836544.0 }, "defaultRanks": { "endpoint": 3, "checkpoint": 86, "checkpointIndependent": 84, "arbitrary": 95, "growthBytesPerToken": 1, "stateBytes": 95 } }, { "id": "LFM2-1.2B", "repo": "LiquidAI/LFM2-1.2B", "name": "LFM2 1.2B", "params": "1.2B", "size": 1.2, "arch": "GQA + 短卷积", "family": "LiquidAI", "layers": 16, "maxContext": 128000, "components": [ { "label": "全长 K + V", "elements": 6144, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "LFM2 gated short-conv 历史", "elements": 40960, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 64, "valueDim": 64, "fullLayers": 6, "convLayers": 10, "convChannels": 2048, "convHistory": 2 }, "notes": [ "conv_L_cache=3 在代码中是 kernel_size；本页保存 K−1=2 槽历史。无额外线性递归矩阵，不把 B/C 门控的瞬时激活当持久状态。" ], "extraSources": [ { "title": "LFM2 短卷积实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/lfm2/modeling_lfm2.py" } ], "configSha256": "55e397257c10a5bb10d4b5bf6013b3cab6d7fb978e7f9533278a9720de6e4556", "configUrl": "https://huggingface.co/LiquidAI/LFM2-1.2B/blob/main/config.json", "cardUrl": "https://huggingface.co/LiquidAI/LFM2-1.2B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "6 层 GQA KV + 10 × 2048 × 2 × 2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 402735104.0, "arbitrary": 3087007744.0, "checkpoint": 407896064.0, "checkpointIndependent": 407896064.0, "globalBytes": 402653184.0, "windowBytes": 0, "stateBytes": 81920, "recurrentBytes": 0, "convBytes": 81920, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 12288.0, "endpointBytesPerToken": 12290.5, "checkpointBytesPerToken": 12448.0, "checkpointIndependentBytesPerToken": 12448.0, "arbitraryBytesPerToken": 94208.0 }, "defaultRanks": { "endpoint": 12, "checkpoint": 1, "checkpointIndependent": 1, "arbitrary": 23, "growthBytesPerToken": 14, "stateBytes": 67 } }, { "id": "LFM2-2.6B", "repo": "LiquidAI/LFM2-2.6B", "name": "LFM2 2.6B", "params": "2.6B", "size": 2.6, "arch": "GQA + 短卷积", "family": "LiquidAI", "layers": 30, "maxContext": 128000, "components": [ { "label": "全长 K + V", "elements": 8192, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "LFM2 gated short-conv 历史", "elements": 90112, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 64, "valueDim": 64, "fullLayers": 8, "convLayers": 22, "convChannels": 2048, "convHistory": 2 }, "notes": [ "conv_L_cache=3 在代码中是 kernel_size；本页保存 K−1=2 槽历史。无额外线性递归矩阵，不把 B/C 门控的瞬时激活当持久状态。" ], "extraSources": [ { "title": "LFM2 短卷积实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/lfm2/modeling_lfm2.py" } ], "configSha256": "68695d4ae8c0db817eec056380741d3e74886c7be310168dfa1f07592653e517", "configUrl": "https://huggingface.co/LiquidAI/LFM2-2.6B/blob/main/config.json", "cardUrl": "https://huggingface.co/LiquidAI/LFM2-2.6B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "8 层 GQA KV + 22 × 2048 × 2 × 2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 537051136.0, "arbitrary": 6442450944.0, "checkpoint": 548405248.0, "checkpointIndependent": 548405248.0, "globalBytes": 536870912.0, "windowBytes": 0, "stateBytes": 180224, "recurrentBytes": 0, "convBytes": 180224, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 16384.0, "endpointBytesPerToken": 16389.5, "checkpointBytesPerToken": 16736.0, "checkpointIndependentBytesPerToken": 16736.0, "arbitraryBytesPerToken": 196608.0 }, "defaultRanks": { "endpoint": 16, "checkpoint": 4, "checkpointIndependent": 4, "arbitrary": 47, "growthBytesPerToken": 17, "stateBytes": 69 } }, { "id": "LFM2-8B-A1B", "repo": "LiquidAI/LFM2-8B-A1B", "name": "LFM2 8B A1B", "params": "8B / A1B", "size": 8, "arch": "GQA + 短卷积", "family": "LiquidAI", "layers": 24, "maxContext": 128000, "components": [ { "label": "全长 K + V", "elements": 6144, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "LFM2 gated short-conv 历史", "elements": 73728, "kind": "fixed", "dtype": "bf16", "role": "state", "ratio": 1, "window": null, "stateClass": "convolution" } ], "geometry": { "kvHeads": 8, "keyDim": 64, "valueDim": 64, "fullLayers": 6, "convLayers": 18, "convChannels": 2048, "convHistory": 2 }, "notes": [ "conv_L_cache=3 在代码中是 kernel_size；本页保存 K−1=2 槽历史。无额外线性递归矩阵，不把 B/C 门控的瞬时激活当持久状态。" ], "extraSources": [ { "title": "LFM2 短卷积实现", "url": "https://github.com/huggingface/transformers/blob/main/src/transformers/models/lfm2/modeling_lfm2.py" } ], "configSha256": "3a63e1ee647926ef3c5eabb7d3a4b5feb38ccece9c624679fa34289692464fa4", "configUrl": "https://huggingface.co/LiquidAI/LFM2-8B-A1B/blob/main/config.json", "cardUrl": "https://huggingface.co/LiquidAI/LFM2-8B-A1B", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "6 层 GQA KV + 18 × 2048 × 2 × 2", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 402800640.0, "arbitrary": 5234491392.0, "checkpoint": 412090368.0, "checkpointIndependent": 412090368.0, "globalBytes": 402653184.0, "windowBytes": 0, "stateBytes": 147456, "recurrentBytes": 0, "convBytes": 147456, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 12288.0, "endpointBytesPerToken": 12292.5, "checkpointBytesPerToken": 12576.0, "checkpointIndependentBytesPerToken": 12576.0, "arbitraryBytesPerToken": 159744.0 }, "defaultRanks": { "endpoint": 13, "checkpoint": 2, "checkpointIndependent": 2, "arbitrary": 39, "growthBytesPerToken": 14, "stateBytes": 68 } }, { "id": "Step-3.5-Flash", "repo": "stepfun-ai/Step-3.5-Flash", "name": "Step 3.5 Flash", "params": "196B / A11B", "size": 196, "arch": "GQA + SWA", "family": "stepfun-ai", "layers": 45, "maxContext": 262144, "components": [ { "label": "全长 K + V", "elements": 24576, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "SWA K + V", "elements": 67584, "kind": "window", "dtype": "selected", "role": "main", "ratio": 1, "window": 512, "stateClass": null } ], "geometry": { "kvHeads": 8, "keyDim": 128, "valueDim": 128, "fullLayers": 12, "windowLayers": 33 }, "notes": [ "layer_types 末尾 3 项属于 MTP，只取前 45 个 backbone block；num_attention_groups 是 KV 头数。" ], "extraSources": [ { "title": "Step 3.5 主干与 KV 分组实现", "url": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/blob/main/modeling_step3p5.py" } ], "configSha256": "794dcd0ea3e925efd9f8bd98745e2246ae73eab58fb0113c3338d7cab619ea3d", "configUrl": "https://huggingface.co/stepfun-ai/Step-3.5-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/stepfun-ai/Step-3.5-Flash", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "2 × 8 × 128 × b × (12T + 33min(T,512))", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 1679818752.0, "arbitrary": 6039797760.0, "checkpoint": 6039797760.0, "checkpointIndependent": 6039797760.0, "globalBytes": 1610612736.0, "windowBytes": 69206016.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 49152.0, "endpointBytesPerToken": 51264.0, "checkpointBytesPerToken": 184320.0, "checkpointIndependentBytesPerToken": 184320.0, "arbitraryBytesPerToken": 184320.0 }, "defaultRanks": { "endpoint": 44, "checkpoint": 63, "checkpointIndependent": 63, "arbitrary": 44, "growthBytesPerToken": 43, "stateBytes": 1 } }, { "id": "MiMo-V2-Flash", "repo": "XiaomiMiMo/MiMo-V2-Flash", "name": "MiMo V2 Flash", "params": "309B / A15B", "size": 309, "arch": "GQA + SWA", "family": "XiaomiMiMo", "layers": 48, "maxContext": 262144, "components": [ { "label": "全局不等维 K + V", "elements": 11520, "kind": "linear", "dtype": "selected", "role": "main", "ratio": 1, "window": null, "stateClass": null }, { "label": "SWA 不等维 K + V", "elements": 99840, "kind": "window", "dtype": "selected", "role": "main", "ratio": 1, "window": 128, "stateClass": null } ], "geometry": { "fullLayers": 9, "windowLayers": 39, "kvHeads": 4, "keyDim": 192, "valueDim": 128, "localKvHeads": 8, "window": 128 }, "notes": [ "全局 4 KV 头、局部 8 KV 头；K=192、V=128，分别相加。attention sink 是学习参数 / bias，不是额外历史 token。" ], "extraSources": [], "configSha256": "3f04b3a6c7edc7062e2501dd83a7726c73a1a0383494c7dd4a8a9ab6ca046d79", "configUrl": "https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash/blob/main/config.json", "cardUrl": "https://huggingface.co/XiaomiMiMo/MiMo-V2-Flash", "evidence": "官方配置 + 状态公式", "tier": "扩充代表", "partial": false, "new": true, "formula": "9 × 4 × (192+128) × b × T + 39 × 8 × (192+128) × b × min(T,128)", "scope": "单序列、单逻辑副本；持久前向状态；压缩 MLA；不含权重、临时激活、投机分支和页对齐", "defaultMetrics": { "endpoint": 780533760.0, "arbitrary": 7298088960.0, "checkpoint": 2390753280.0, "checkpointIndependent": 2390753280.0, "globalBytes": 754974720.0, "windowBytes": 25559040.0, "stateBytes": 0, "recurrentBytes": 0, "convBytes": 0, "otherStateBytes": 0, "checkpointCount": 64, "growthBytesPerToken": 23040.0, "endpointBytesPerToken": 23820.0, "checkpointBytesPerToken": 72960.0, "checkpointIndependentBytesPerToken": 72960.0, "arbitraryBytesPerToken": 222720.0 }, "defaultRanks": { "endpoint": 23, "checkpoint": 26, "checkpointIndependent": 26, "arbitrary": 51, "growthBytesPerToken": 25, "stateBytes": 1 } } ], "methodology": { "units": "KiB=1024 B, MiB=2^20 B, GiB=2^30 B；b=KV 精度/8；递归精度独立选择；卷积 BF16，压缩器 FP32，ID int64。", "endpoint": "G(T)+W(T)+S：在长度 T 的同一个边界保存全部持久状态，可续算后续输入。", "arbitrary": "G(T)+Σa_window*T+T*S：每个 token 保存独立状态快照；局部 KV 历史保留一份，允许任意 token 边界不重放恢复。", "checkpoints": "E={P,2P,...,floor(T/P)P}∪{T}，去重；G(T)+ΣW(E)+|E|S。G 只保留一份；SWA 可用区间并集去重。", "conventions": [ "S 是完整的固定形状状态快照，包含卷积和尾块；不做递归状态 delta 编码，也不跨快照去重卷积 / 压缩器 / n-gram 内部历史。", "有限 SWA 按配置 W 个历史槽；短卷积按最小 K−1 历史规范化。逻辑载荷不是张量分配器实际显存。", "每层恢复边界必须一致；token 哈希、模型版本、精度、位置语义一致才能命中。", "checkpoint 不含持续解码的额外工作副本，恢复时可能增加一份可变状态。整段前缀末端快照与工作状态在逻辑大小上相同。", "保存状态不等于保存整个进程：不含调度器、RNG、CUDA Graph、采样器、last logits。若要求缓存命中后不跑最后一个 token 就直接采样，需要另存最后 logits，或最后 hidden 并重算 LM head。", "所有计算是配置 / 代码推导，未加载模型、未生成实际状态 tensor、未测 GPU 显存；非能力/SOTA 分数排行。" ] }, "sourceSnapshots": [ { "file": "HuggingFaceTB__SmolLM2-1.7B-Instruct.json", "sha256": "994f50b16abb4ae00880baefe03c10260b5bd608d2bf586f7056ca05a534feea" }, { "file": "HuggingFaceTB__SmolLM3-3B.json", "sha256": "c72b1031274ff4626e434d0019e88e95a767460135db9ee492eb80652b786af1" }, { "file": "LiquidAI__LFM2-1.2B.json", "sha256": "55e397257c10a5bb10d4b5bf6013b3cab6d7fb978e7f9533278a9720de6e4556" }, { "file": "LiquidAI__LFM2-2.6B.json", "sha256": "68695d4ae8c0db817eec056380741d3e74886c7be310168dfa1f07592653e517" }, { "file": "LiquidAI__LFM2-8B-A1B.json", "sha256": "3a63e1ee647926ef3c5eabb7d3a4b5feb38ccece9c624679fa34289692464fa4" }, { "file": "MiniMaxAI__MiniMax-M2.5.json", "sha256": "3372a31ecbdba614b309cfa5b3dfec33ceea3c580cb9612ff35345d92ced8618" }, { "file": "MiniMaxAI__MiniMax-M3.1-Flash.json", "sha256": "df6a3668e2de5b8dc0e883e3767495ad17af7e4d69fabe9683088be015fda9ab" }, { "file": "MiniMaxAI__MiniMax-M3.1-Flash__README.md", "sha256": "c49a7cca92321066bc20e25ce720260242dcb112baf1d09c7c213899e39d2151" }, { "file": "Qwen__Qwen2.5-1.5B-Instruct.json", "sha256": "98d2ff8cc47488d08a2b0b3acf4eb99ef210779b42bd48605f6b8e36acdbf670" }, { "file": "Qwen__Qwen2.5-14B-Instruct.json", "sha256": "0f2085dbbe2ee251bd6a6a0797d84a6ce34436044d629aa3cba793b43d311a9e" }, { "file": "Qwen__Qwen2.5-32B-Instruct.json", "sha256": "9c6772f138ef9e5b3d1c18f2c87e451bbc01f5f1a4eabb36f9bf4f53829b903e" }, { "file": "Qwen__Qwen2.5-3B-Instruct.json", "sha256": "eed00b17e22553979d090fa492e587e92885e328914c8e0b0b78f0a0d3576b3b" }, { "file": "Qwen__Qwen2.5-72B-Instruct.json", "sha256": "14ca217334fe0fd10148413592d68c99eeb33431ed89c1afa130fee560be2a29" }, { "file": "Qwen__Qwen2.5-7B-Instruct.json", "sha256": "7463bb0ea78315365e6c6b74de4e73bbcc8359dfb0c5a737584e077d42c0b03c" }, { "file": "Qwen__Qwen3-1.7B.json", "sha256": "1ddb5b89ebc90dcb417a45c213d818577e65976454d29385c8f6140771d95197" }, { "file": "Qwen__Qwen3-14B.json", "sha256": "e73c3664ca09b10a673fef0c22e8a6b456201d49bd4713c9691f775720e8857a" }, { "file": "Qwen__Qwen3-235B-A22B-Instruct-2507.json", "sha256": "ad59a339665ee1c68cbff06ac72c9e59f5183ebfaf3698c98737e003a870e9e2" }, { "file": "Qwen__Qwen3-30B-A3B.json", "sha256": "2850ddb3bf7aecad20b611e2d44f3077fc8193f4827c93beddd4c02ad63c2297" }, { "file": "Qwen__Qwen3-32B.json", "sha256": "97e295b63283935788fac5e4f8860862a56d4089538cafc93f0431f2ebe483bb" }, { "file": "Qwen__Qwen3-4B.json", "sha256": "8ba006f74fecfaaeb392872a60f4a480e7ec9860153d2e1b769ec81f9a147f8a" }, { "file": "Qwen__Qwen3-8B.json", "sha256": "f7c4eadfbbf522470667b797a3c89be2524832d2d599797248dc304fff447c30" }, { "file": "Qwen__Qwen3-Coder-480B-A35B-Instruct.json", "sha256": "653230e2dd1c41bd1125d046f00259ed98d6294c75c63969f8890c6e139af7de" }, { "file": "Qwen__Qwen3-Coder-Next.json", "sha256": "a7b8098d3b05777f12bb5677a26bf1240a1bb09def1b06b29e6be86cae2e84f8" }, { "file": "Qwen__Qwen3-Next-80B-A3B-Instruct.json", "sha256": "2d483c7cabad7c8704478ed4038fa7e7b2eff840bc00a118eccbe38e2b488303" }, { "file": "Qwen__Qwen3.5-0.8B.json", "sha256": "b90b86f35c8e6925ef74ee04d0e758f0a845c83a42089ad82bbaa948de9b4204" }, { "file": "Qwen__Qwen3.5-122B-A10B.json", "sha256": "af07d0423658865afb52bd0d6fd3c2cc45d988425b76193467fd19129655cecd" }, { "file": "Qwen__Qwen3.5-27B.json", "sha256": "f8d190c5b89c1521220f935d2567a587d6e291ed69066a45a106560b05a2174c" }, { "file": "Qwen__Qwen3.5-2B.json", "sha256": "ed1c1723241f23f7f4e23430759cbd7dcfb4103cbdfe052bfe7626b57c2615b4" }, { "file": "Qwen__Qwen3.5-35B-A3B.json", "sha256": "5e4d7f74fec2f360eb9cfbfcd6ec0c4c76e684d3a11caaed259d9fd9bfbc7944" }, { "file": "Qwen__Qwen3.5-397B-A17B.json", "sha256": "3ae7fa89c2f7d1354096418ddaf1331e9e0898a8ca11e804fb6dbaa087efb7da" }, { "file": "Qwen__Qwen3.5-4B.json", "sha256": "ddc63e1c717afa86c865bb5e01313d89d72bb53b97ad4a8a03ba8510c0621670" }, { "file": "Qwen__Qwen3.5-9B.json", "sha256": "d0883072e01861ed0b2d47be3c16c36a8e81c224c7ffaa310c6558fb3f932b05" }, { "file": "Qwen__Qwen3.6-27B.json", "sha256": "69db4eb7196bc8190813231b3018ca05d8c2e3abc7b1af19d55c157af44a9d9c" }, { "file": "Qwen__Qwen3.6-35B-A3B.json", "sha256": "93a4693fa9d8392fbfccd4b3c9873f4bfdcb14fdede978b123d07d19675efe99" }, { "file": "Qwen__Qwen3.8-2.4T-A95B.json", "sha256": "4e3819548967e319ab435d044a3a331dbe3b078590ce822e9d74b79430533987" }, { "file": "Qwen__Qwen3.8-Flash-Next.json", "sha256": "889658f2508e8c61d409b02e70e0d78d8d4452ec65aaafbe129805d213d2e74b" }, { "file": "XiaomiMiMo__MiMo-V2-Flash.json", "sha256": "3f04b3a6c7edc7062e2501dd83a7726c73a1a0383494c7dd4a8a9ab6ca046d79" }, { "file": "XiaomiMiMo__MiMo-V2.6-Pro-RL__README.md", "sha256": "6a404a7763105f70d5b259b06306562d192735a905e76fab1609ecad79fdbd0a" }, { "file": "Zyphra__Zamba2-1.2B-instruct.json", "sha256": "3cc6e1fd1266a062390f50cd5e13ae36e2f36ada8dbb3443c284581af5eb355a" }, { "file": "Zyphra__Zamba2-2.7B-instruct.json", "sha256": "0497ca3001366ed8eaef1030c84f856fd864f6f5e2ba86cef9ee9b211807df02" }, { "file": "Zyphra__Zamba2-7B-instruct.json", "sha256": "51d90f27707cd78e64184c5f5295f5b2dcc967c75f4469d522b198a0f245f5e8" }, { "file": "allenai__OLMo-2-0325-32B-Instruct.json", "sha256": "e977b63b4196481a109089de5a4d4479eb18df70297d9c3e18d865cf5b098e51" }, { "file": "allenai__OLMo-2-0425-1B.json", "sha256": "3dc581384d61b3366fc34a95d12c20e595014ea62352469c17c3e3a77f32882a" }, { "file": "allenai__OLMo-2-1124-13B-Instruct.json", "sha256": "10b1de896314ca2dbf5aeda0035113dbeaa34e086f2a29bb4944596a15f4861e" }, { "file": "allenai__OLMo-2-1124-7B-Instruct.json", "sha256": "ff8cc8709a229515676797ab6f343a09391041c9a8fbbc78bfec5be4c2e3664e" }, { "file": "allenai__Olmo-3-7B-Instruct.json", "sha256": "6282ae812f9e494bd212e9a3233d68dc4d5f1e2569642423d2b79a3e80373ef8" }, { "file": "allenai__Olmo-3.1-32B-Instruct.json", "sha256": "d47a120cb2669cd46482d5c2da9488903b67759ca167cd85c1127b2e05755f95" }, { "file": "baidu__ERNIE-4.5-21B-A3B-PT.json", "sha256": "7383c7387390746c1fa0327f8cab3d5c6b078a60bbf0841d4f42277eece5a5f6" }, { "file": "baidu__ERNIE-4.5-300B-A47B-PT.json", "sha256": "607db524c1233d2687e10df8cb049bcc781a5ba232ee3476553450cf9f5bba75" }, { "file": "deepseek-ai__DeepSeek-R1.json", "sha256": "79ddea672a62e95d3f0be27be434375538e1988975971cffcf87c96c1de84a65" }, { "file": "deepseek-ai__DeepSeek-V2-Lite.json", "sha256": "f346286b0f1c8b044252fd54cb4fa78b9fab6472a6e8bebb9edfe03d414ea03d" }, { "file": "deepseek-ai__DeepSeek-V2.5.json", "sha256": "4cb37d1e07da958c0367cc9697cec37ae0a80ee9a7c922c37232d798fb311e3e" }, { "file": "deepseek-ai__DeepSeek-V2.json", "sha256": "36f572d2909bd108d73a814c585fd61afd56266960a13eb67c713e229d3435d1" }, { "file": "deepseek-ai__DeepSeek-V3.2.json", "sha256": "c7fa8b191e9936d8e6a57d864baab82b792fae16a116416cdd3a75ba76bc5af1" }, { "file": "deepseek-ai__DeepSeek-V3.json", "sha256": "cbf0b95dc614de208a109bb5fd4e7eed11385e9c68411d2c17db5319443035d9" }, { "file": "deepseek-ai__DeepSeek-V4-Flash.json", "sha256": "b628e63398a645abc711d92207f8737dd8140f7a4ef1e0a5b3616019e0ddd818" }, { "file": "deepseek-ai__DeepSeek-V4-Pro.json", "sha256": "5fe4568daee51c208cb8a79538eaeda090ae011ade1dee2c386aa95f569c810e" }, { "file": "deepseek-ai__DeepSeek-V4-Pro__README.md", "sha256": "c4d714818a4d3333542edc7d38ea065825a0cf7aa8fea3605bbd1d1c18e4a610" }, { "file": "deepseek-ai__DeepSeek-V4.1-Flash.json", "sha256": "8be45ce0476004a3f529fd896115a4a2e800a129ad2d3ec05b16050f52e21879" }, { "file": "deepseek-ai__DeepSeek-V4.1-Flash__README.md", "sha256": "347c9db4e5506acb531cbc3b724407ab88e9af8781679152f0823d7bac16d251" }, { "file": "deepseek-impl.py", "sha256": "314e533c1606278d8294334edf13a23f082f1687a06f5e9bdf92a4958775ea83" }, { "file": "deepseek-v32-model.py", "sha256": "bfe5b89186b579910b6d59fa7e0cf1f7a75ceb4ec5d0f5b54f0ccbe77cdb6a63" }, { "file": "deepseek-v4-model.py", "sha256": "ce962f1face79d4f633d36436576214057a7e11443c9789935e1deb5c6cd1d71" }, { "file": "deepseek-v41-model.py", "sha256": "e5decd76fe70fd56ef2f10124d245c6f64fb6a0003cc9068a1d3182a168b8d2d" }, { "file": "dsa-kpool.py", "sha256": "db386335724accbbbeaaf5396be8af3b22fa71b5a9cbeb57f5b4198e0c9a0efc" }, { "file": "falcon-h1.py", "sha256": "b4e2180d8262518271a1e3c7f11533e251cb6ce952b990b91efd197ff700ab9e" }, { "file": "gemma4-config.py", "sha256": "961665805532efd0829aa5665c6b9d2c12ac43a0ff91b75eadca421d08ad8ce0" }, { "file": "glm-dsa-hf.py", "sha256": "226f5a24b6aa15eca90bcc74e6bf1e2e887687ccfeb0d1ae28da4bf4c216038f" }, { "file": "glm-dsa-model.py", "sha256": "2ecfafed54e99428fa7cac102a5e25de0945872c88d4d77c438b9632c5abc447" }, { "file": "glm-next.py", "sha256": "82868ef8dc9d803aec2c4c24d412125b894d6518ee602e9780f64b899171c5f4" }, { "file": "google__gemma-4-12B.json", "sha256": "14f38c5492ffc9cbcdf808647ca0c025bb5b9b4eb737526347134d500ace6098" }, { "file": "google__gemma-4-26B-A4B.json", "sha256": "0364256c5de7863a22b5dd666043d542d42b681c02fd83eb67dc7b483d859e96" }, { "file": "google__gemma-4-31B.json", "sha256": "6a81841cad2b6ba06841e23c6afb5f0a27827bc12c64328ffa6338831a21267e" }, { "file": "google__gemma-4-E2B.json", "sha256": "e5faef0dd1a8f2437f6010721146b85433eaa90e679ef011e803c7ffefae73b8" }, { "file": "google__gemma-4-E2B__README.md", "sha256": "4c4e89a23e46a5703148b4de82eafb5b650abc8f50ff616c090fb659602eb4a6" }, { "file": "google__gemma-4-E4B.json", "sha256": "f27a045f32c39fb9cd930204920de6b0962810cf09929a8810901ffaec780f20" }, { "file": "google__gemma-4-E4B__README.md", "sha256": "23ba22e9aadcf727659e511de593c6d365d87fc0f3a66dd297769aef591f5d38" }, { "file": "granite-hybrid.py", "sha256": "7a8fa98d08fccad6b5e3328a15fe84f2444fd0e36ada392e9b3996a5cb7a218f" }, { "file": "granite-small-readme.md", "sha256": "d8afa48cec7096978d19fdf19fd9012b9505f428c5d64c231fde63241be176ea" }, { "file": "ibm-granite__granite-4.0-h-1b.json", "sha256": "aafbdabf9b0ff18ab18e6d22b867b1ba006305fd666a8d9f579de7c4a98ca0b8" }, { "file": "ibm-granite__granite-4.0-h-micro.json", "sha256": "bd828b00edcc25f93d6fa55210acebcd058054c4d5b5fa6c5146b0a0d18172ee" }, { "file": "ibm-granite__granite-4.0-h-small.json", "sha256": "8616e9f0b30e6fac9696f7c1e1dbd08f1a850ac4af0de6353f7d6009043702ae" }, { "file": "ibm-granite__granite-4.0-h-tiny.json", "sha256": "bda8fd574ace7d968d82397f59ea6b9a702a077bbeab279a65b9dad7386a82c6" }, { "file": "ibm-granite__granite-4.2-30b.json", "sha256": "82c834ac6bf6a2ecbf9801b49c3fb927008df24ea415c35e931b6091cab21277" }, { "file": "ibm-granite__granite-4.2-3b.json", "sha256": "e4f44b9c6924a283f1d1d36057a89f1a04404bf1561fc2c0b64cf98180e72915" }, { "file": "ibm-granite__granite-4.2-8b.json", "sha256": "85611f4e34633d4e148e6a5f64bc2d3a23ebbc2b014bc87fbda764eecf261cd9" }, { "file": "internlm__internlm2_5-20b-chat.json", "sha256": "0c2a8521c46627f3189d18f804b3709b179b211d43b4527120e0173e0ce4b8cf" }, { "file": "internlm__internlm3-8b-instruct.json", "sha256": "e1d88b33517b569d39ccb6730548f6eea9adb5dac892d5db74b78f670c0a7471" }, { "file": "jamba.py", "sha256": "e78e4d694740959391848456b0dc1a97246b9d3c3d7b311d3f28c717401a57af" }, { "file": "kimi-config.py", "sha256": "1e50e8715772e83c1dd2316c33e66d00b8d51c86ab7e3968aab7522a5290dd2b" }, { "file": "kimi-k3-model.py", "sha256": "9e3564c70ac21854ce5a090cc946c5dc76b70d1050ef50840449181a20fff44a" }, { "file": "kimi-linear.py", "sha256": "d79b365e37378881b9f1585007a56e236ca27a414920943cb85d1dacb75dda99" }, { "file": "lfm2.py", "sha256": "c4940553e30be672ec8bea4290138f2e4c98944be96c9361f4154f4550e3aebc" }, { "file": "llama-skus.py", "sha256": "a71a8ca23c8daf3d1daec02c13e9916c36e315c5056b56ed20b18df98d10c486" }, { "file": "mamba-config.py", "sha256": "2a72c1686f775b56547e39ca4406ba10148d12fd7a791c57ce2ba85126010fcd" }, { "file": "mamba-utils.py", "sha256": "659d941b57b6960f1144e7c6a6e0449ea630b273ea69cb4f89312c10e0a8ff11" }, { "file": "mamba2-module.py", "sha256": "605e4439ff0baec8d8acaf4a191d9f0570eea9900065a065909124c472b08707" }, { "file": "memory-pool.py", "sha256": "b1a49c3d476e40fd036d38300d56eb6a6530297510a4936f646e066831cf1348" }, { "file": "microsoft__Phi-4-mini-instruct.json", "sha256": "ac65d86061d3d0d704ee2511fd0eb8713ef19eb6eedba17c3080a4165d5b933b" }, { "file": "microsoft__phi-4.json", "sha256": "07eedad2c48798b6e3728e4a1b75e0e092019a375ba725a40e40c78d13664045" }, { "file": "mimo-model.py", "sha256": "a8c3cb3aae473bcc15f023010547c919f15eba6546e6ed7efb61a8937b12f3ad" }, { "file": "mimo.json", "sha256": "a9dc00e2a0c2c172efd198f42e81e41c95472021b57c65366c06587d44038b62" }, { "file": "minimax-model.py", "sha256": "767aadfed5133848b8f3a31aa39219b700eb39f7e45e6b256c56be37d42edffc" }, { "file": "mistralai__Ministral-3-14B-Instruct-2512.json", "sha256": "fc51295bdc0b39c015c9c3ca5d649de183eae73a23b8cbab1a0e055be8d5fa04" }, { "file": "mistralai__Ministral-3-3B-Instruct-2512.json", "sha256": "a9c3658d7bb45ab5be77e660d483a9d3f8a3d54bda2c69a1c576bdfefbc16be6" }, { "file": "mistralai__Ministral-3-8B-Instruct-2512.json", "sha256": "196c38570bb4bc8168a59cc0189c9a9b9974be8dcca6b46dfc109dbd250c01aa" }, { "file": "mistralai__Mistral-7B-Instruct-v0.3.json", "sha256": "affafc6478ec0fd07a32f0ca57aa2fc57743f4d17d6730f86a96ac24d1507f99" }, { "file": "mistralai__Mistral-Large-Instruct-2411.json", "sha256": "2c379704c86ffd8273b1aa9beae0ce20c4c36f47a0c86708efb5aa46c0ac7b74" }, { "file": "mistralai__Mistral-Small-3.2-24B-Instruct-2506.json", "sha256": "01ab910a5dda7995709cc355d094eabb8094b78d49240cd167188606c3ff5edb" }, { "file": "mistralai__Mixtral-8x22B-Instruct-v0.1.json", "sha256": "9c4a6138d84029ab666943613e3d5844d2ea8fd6149f44f77188c62e2915e0f5" }, { "file": "mistralai__Mixtral-8x7B-Instruct-v0.1.json", "sha256": "9d56d04b36d0fd12ff54ae4c5bac769cc176e254e64ff71144614b6318b40793" }, { "file": "moonshotai__Kimi-K2.5.json", "sha256": "acd5bb01a16f64b309599cd6ed196be056f613c99d6bc9300692b82cd10882f6" }, { "file": "moonshotai__Kimi-K2.6.json", "sha256": "85825ca6e18cbe539eb83ee09eedfb3f4222265929f06e9f535a6d9364f55899" }, { "file": "moonshotai__Kimi-K3.json", "sha256": "9710e121a58d03ac92c8d6da287a19541994319afbbe6d6202af001ffd379213" }, { "file": "moonshotai__Kimi-K3__README.md", "sha256": "57de265b5842dfa465c6e73b368b0e15a89b8793b5450528dad577da202cc6fe" }, { "file": "moonshotai__Kimi-Linear-48B-A3B-Instruct.json", "sha256": "a6ac3c2c4b5aa72370f9727f49ffa4432715d20061889acdb37c688be853096e" }, { "file": "nemotron-h-config.py", "sha256": "0fc818c10506c91bd02df5a605f49cb0704b5498954f46dbde2d63999ae36c3d" }, { "file": "nemotron-h.py", "sha256": "e1cb5fc02e887983f0a445bf4c1a2604453b2cb2db4624c7004dcf663bbb1b6e" }, { "file": "nemotron-nano4-config.py", "sha256": "07fa66e5b3da7e6a71c1a263e3dd68da11c8afa9178b47c49510ba628746fcff" }, { "file": "nvidia__NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.json", "sha256": "dd9fa380ac107b0477db5a26108db9febe6378e7bb3966a107944853ec4f76f8" }, { "file": "nvidia__NVIDIA-Nemotron-3-Nano-4B-FP8.json", "sha256": "1a0a0703a73600ff4b1042dde35cb2137c7e4d76dadc5a88f6f436f2323aa3e9" }, { "file": "nvidia__NVIDIA-Nemotron-3-Super-120B-A12B-BF16.json", "sha256": "699f34f0fc645d29ebffa5767fb59e6ae6ec98e3a4605485eb9913256d0df7e6" }, { "file": "openai__gpt-oss-120b.json", "sha256": "933aeb666a3fd851133ddd7686414f369bc564c4185fb5704416550879f10566" }, { "file": "openai__gpt-oss-20b.json", "sha256": "3a2a26ded679375b7928ddeca59764df7cea83220c1961035f6d6e232659e9ce" }, { "file": "qsa-glue.py", "sha256": "025fd7a4e72dd6948d692282dbe56f73c1c1d2b66e40d3b9e83b3a196f17115e" }, { "file": "qsa-indexer.py", "sha256": "4b31def7d5351815e77ba0ecc13e6425919bad2b759d68d0598e7f1676a7d335" }, { "file": "qsa-pool.py", "sha256": "3e1cab68bd4de7be176249b8fb68e3c1c63bd53030c8e43f0df38e5dd57309fe" }, { "file": "qwen-next.py", "sha256": "33fec8a53f90cc2932db7a732f0ceab4daa80d2ff79e766d500fe833833624b7" }, { "file": "qwen38-report.md", "sha256": "7690c680c2d6eb284179ef5dd4db63260d248531cae9087b09b2c2200230f331" }, { "file": "sglang-kimi.py", "sha256": "b45907fe80168e8c5706be165c313d7590e3c63ae6f4133c160959c63be3fa04" }, { "file": "state-spaces__mamba2-1.3b.json", "sha256": "204f2030f433f61e53fd9c3639662c6cf231cd390071f6ea38b0890b0e3c04f0" }, { "file": "state-spaces__mamba2-2.7b.json", "sha256": "0955f133a3eeb902ed11e1a65d0be2934e33a8e1871b6a9685414fa8955e6c86" }, { "file": "step35.py", "sha256": "b17c4483f30b4f67c8ac0194aaa652ce7a9e6b8efd9ffe08445de4e9be538f87" }, { "file": "stepfun-ai__Step-3.5-Flash.json", "sha256": "794dcd0ea3e925efd9f8bd98745e2246ae73eab58fb0113c3338d7cab619ea3d" }, { "file": "tencent__Hunyuan-A13B-Instruct.json", "sha256": "5fb4af8fa99a28c0bfd1ae009b9488acb2c4365956ffab472c38ae2c1a3d378e" }, { "file": "tiiuae__Falcon-H1-1.5B-Instruct.json", "sha256": "2473f5e02854a2833df973277e6f109d94344e7860fbc78d5ff8b31d08dddaeb" }, { "file": "tiiuae__Falcon-H1-34B-Instruct.json", "sha256": "156d23cf8c31a19d1951fb530ce703441b8dba7002af7f6185d26a9d191b1a2f" }, { "file": "tiiuae__Falcon-H1-3B-Instruct.json", "sha256": "f6faecc746e9bd9450747e36e69f802f39551b3eb74fbbf9bef5a7afddb8c12f" }, { "file": "tiiuae__Falcon-H1-7B-Instruct.json", "sha256": "6a91400ff4652977d914fab8c495716cd6020f017b578dde89cc9744e4199607" }, { "file": "transformers-gemma4.py", "sha256": "edb175aa24e0482f7e5eea2794e650bc6e41268d58d831a66caefd5c08b6fe4c" }, { "file": "transformers-qwen35.py", "sha256": "4463b8a108f678c166755101d95f188179f54143305953d78147de0389b80ebc" }, { "file": "v41-engram.py", "sha256": "11f35ecbead8150c35aa002b3d180ef290b05a25afe883a11884f94d476d3897" }, { "file": "zai-org__GLM-4.5-Air.json", "sha256": "e59f21d7be4b4b917d80bb3a2d1211356d29cf3aeff93372523daf7d0ffb839d" }, { "file": "zai-org__GLM-4.5.json", "sha256": "a5dbd07371a8a51b0013a63a7a70543a7094af1a9598812b537133e6da78dda7" }, { "file": "zai-org__GLM-4.7-Flash.json", "sha256": "dc9b97c7c9bed726a2e6939da4234d5c43abb3edec8812068c9a1af1dbc13acb" }, { "file": "zai-org__GLM-4.7.json", "sha256": "f713924a536ac1965d8af2a3144d3a0ce0acae94f0fbfd7407240cbafcb08c53" }, { "file": "zai-org__GLM-5.3-Flash.json", "sha256": "bb8f01c42cb92a52ca72e65afb4d5bd8d11aef083cd210e8de25dfb904f23e9f" }, { "file": "zai-org__GLM-5.3-Flash__README.md", "sha256": "e6c5d41400e28f5a89418a0d15b09b9c113314a7553d40d7d4a8a72fcdb87d02" }, { "file": "zai-org__GLM-5.3.json", "sha256": "3ac72612095574542f7fff847ada8e59d9199dd8af44bdf625d7e02615572e69" }, { "file": "zai-org__GLM-5.3__README.md", "sha256": "ed1c0a4563c437a32f8953d637a1db9bd831de1b24bd3062a5ace9e04340b1cf" }, { "file": "zai-org__GLM-5.json", "sha256": "308d3ceebe7bd9c2246d20550d5b793098dd03838d1e504f8236063ae152bcae" }, { "file": "zamba2.py", "sha256": "2b5b7828f24baae272c9e9d34f97dc5e6d81923f2452ae797c9ca99e9283923d" } ] }