{ "models": { "deepseek/v4.1-flash-1017": { "name": "DeepSeek V4.1 Flash 1017", "short": "DS V4.1 Flash", "arch": "MoE · encoder/decoder 各 20 层（YOCO 类）· 稀疏注意力 · Engram", "totalB": 552, "sideB": 196.6, "activeP": 8, "activeD": 16, "layers": 40, "hidden": 5120, "streams": 4, "nExp": 384, "topk": 6, "expParams": 35389440, "denseLayers": [], "denseFfnParams": 0, "nonExpParams": 170000000, "embedParams": 661913600, "headParams": 661913600, "formats": { "nvfp4": { "bpp": 0.5625, "compute": "fp4", "label": "NVFP4" }, "mxfp4": { "bpp": 0.53125, "compute": "fp4", "label": "MXFP4（官方）" } }, "mtpGB": { "nvfp4": 7.2, "mxfp4": 7.2 }, "side": { "name": "Engram", "gb": 104.4, "layers": [ 1, 14 ], "rowsPerTokenPerLayer": 8, "rowBytes": 136 }, "kvFp4": true, "kv": { "perToken": { "fp4": 890, "fp8": 1700, "bf16": 3200 }, "kvHeads": 0, "stateLayers": 0, "stateMB": 0 }, "encoderLayers": 20, "tail": 128, "baseUs": 40, "spec": { "name": "DSpark", "verify": 5, "accept": 2.8 }, "src": "config.json（nvidia/DeepSeek-V4.1-Flash-NVFP4）、vLLM recipe" }, "zai/glm-5.3": { "name": "GLM-5.3", "short": "GLM-5.3", "arch": "MoE · 78 层（前 3 层 dense）· DSA 稀疏注意力 · MLA", "totalB": 753, "sideB": 0, "activeP": 40, "activeD": 40, "layers": 78, "hidden": 6144, "streams": 1, "nExp": 256, "topk": 8, "expParams": 37748736, "denseLayers": [ 0, 1, 2 ], "denseFfnParams": 226492416, "nonExpParams": 250000000, "embedParams": 951582720, "headParams": 951582720, "formats": { "fp8": { "bpp": 1, "compute": "fp8", "label": "FP8（官方）" }, "nvfp4": { "bpp": 0.5625, "compute": "fp4", "label": "NVFP4（社区，估）" } }, "mtpGB": { "fp8": 10.5, "nvfp4": 5.9 }, "side": null, "kv": { "perToken": { "fp4": 30900, "fp8": 54900, "bf16": 109800 }, "kvHeads": 0, "stateLayers": 0, "stateMB": 0 }, "encoderLayers": 0, "tail": 0, "baseUs": 30, "spec": { "name": "MTP", "verify": 2, "accept": 1.7 }, "src": "AWS/NVIDIA 模型卡；moe_intermediate=2048、注意力参数按 GLM-5 估" }, "qwen/3.8-flash-next": { "name": "Qwen3.8 Flash Next", "short": "Qwen3.8 FN", "arch": "MoE · 48 层（36 Gated DeltaNet : 12 稀疏注意力）· N-gram 表", "totalB": 74, "sideB": 51, "activeP": 6, "activeD": 6, "layers": 48, "hidden": 2560, "streams": 1, "nExp": 512, "topk": 10, "expParams": 2703360, "denseLayers": [], "denseFfnParams": 0, "nonExpParams": 63000000, "embedParams": 388956160, "headParams": 388956160, "formats": { "nvfp4": { "bpp": 0.5625, "compute": "fp4", "label": "NVFP4" }, "fp8": { "bpp": 1, "compute": "fp8", "label": "FP8" } }, "mtpGB": { "nvfp4": 2.3, "fp8": 4 }, "side": { "name": "N-gram 表", "gb": 51, "layers": [ 0 ], "rowsPerTokenPerLayer": 4, "rowBytes": 2560 }, "kv": { "perToken": { "fp4": 6900, "fp8": 12300, "bf16": 24600 }, "kvHeads": 2, "stateLayers": 36, "stateMB": 0.5 }, "encoderLayers": 0, "tail": 0, "baseUs": 22, "spec": { "name": "MTP", "verify": 3, "accept": 2.2 }, "src": "公开报道（125B 总参，含 51B N-gram、4B MTP）；专家维度和 KV 结构为估算" }, "moonshot/kimi-k3": { "name": "Kimi K3", "short": "Kimi K3", "arch": "MoE · 93 层 · KDA 线性注意力 + MLA 混合", "totalB": 2800, "sideB": 0, "activeP": 104, "activeD": 104, "layers": 93, "hidden": 7168, "streams": 1, "nExp": 896, "topk": 16, "expParams": 33030144, "denseLayers": [ 0 ], "denseFfnParams": 396361728, "nonExpParams": 600000000, "embedParams": 1174405120, "headParams": 1174405120, "formats": { "nvfp4": { "bpp": 0.5625, "compute": "fp4", "label": "NVFP4（NVIDIA）" }, "mxfp4": { "bpp": 0.53125, "compute": "fp4", "label": "MXFP4（官方）" } }, "mtpGB": { "nvfp4": 10, "mxfp4": 10 }, "side": null, "kv": { "perToken": { "fp4": 7450, "fp8": 13250, "bf16": 26500 }, "kvHeads": 0, "stateLayers": 70, "stateMB": 2 }, "encoderLayers": 0, "tail": 0, "baseUs": 35, "spec": { "name": "DSpark", "verify": 4, "accept": 2.4 }, "src": "模型卡（2.8T 总参、104B 激活、93 层、896 专家）；隐藏维度和 MLA 层数为估算" } }, "gpus": { "rtxpro5000-72g": { "name": "RTX PRO 5000 72GB", "mem": 72, "bw": 1.344, "fp4": 1, "fp8": 0.5, "pcie": 50, "price": 170000, "tdp": 300 }, "rtxpro5000-48g": { "name": "RTX PRO 5000 48GB", "mem": 48, "bw": 1.344, "fp4": 1, "fp8": 0.5, "pcie": 50, "price": 150000, "tdp": 300 }, "rtxpro6000d-84g": { "name": "RTX PRO 6000D 84GB", "mem": 84, "bw": 1.4, "fp4": 1.5, "fp8": 0.75, "pcie": 50, "price": 180000, "tdp": 600 }, "rtxpro6000-96g": { "name": "RTX PRO 6000 96GB", "mem": 96, "bw": 1.792, "fp4": 2, "fp8": 1, "pcie": 50, "price": 220000, "tdp": 600 }, "dgx-spark": { "name": "DGX Spark (GB10)", "mem": 128, "bw": 0.273, "fp4": 0.5, "fp8": 0.25, "unified": true, "pcie": 0, "price": 35000, "tdp": 240 }, "h200-141g": { "name": "H200 141GB", "mem": 141, "bw": 4.8, "fp4": 0, "fp8": 1.98, "pcie": 50, "price": 300000, "tdp": 700 }, "b300-288g": { "name": "B300 288GB", "mem": 288, "bw": 8, "fp4": 13.5, "fp8": 4.5, "pcie": 100, "price": 600000, "tdp": 1100 } }, "links": { "nvlink5": { "name": "NVLink5", "short": "NVLink5", "bw": 900, "lat": 2, "kind": "nvlink" }, "pcie5-p2p": { "name": "PCIe5 同CPU", "short": "PCIe5", "bw": 50, "lat": 3, "kind": "pcie" }, "pcie5-xgmi": { "name": "PCIe5 跨CPU(xGMI)", "short": "xGMI", "bw": 30, "lat": 6, "kind": "pcie" }, "pcie4-p2p": { "name": "PCIe4 P2P", "short": "PCIe4", "bw": 25, "lat": 4, "kind": "pcie" }, "eth-100g": { "name": "ETH 100G", "short": "ETH 100G", "bw": 11.5, "lat": 10, "kind": "eth" }, "eth-200g": { "name": "ETH 200G", "short": "ETH 200G", "bw": 23, "lat": 10, "kind": "eth" }, "eth-400g": { "name": "ETH 400G", "short": "ETH 400G", "bw": 46, "lat": 10, "kind": "eth" }, "eth-800g": { "name": "ETH 800G", "short": "ETH 800G", "bw": 92, "lat": 10, "kind": "eth" } }, "modes": { "agg": { "name": "混合部署", "desc": "Prefill 与 Decode 同一组卡，chunked prefill 混批" }, "pd": { "name": "PD 分离", "desc": "P 放全部层，KV 传给 D" }, "pd-asym": { "name": "非对称 PD", "desc": "P 只放 encoder 层，D 补算末尾 token", "yoco": true }, "layer-split": { "name": "按层拆", "desc": "A 组跑 encoder（含全部 Prefill），B 组跑 decoder，Decode 跨组流水", "yoco": true } }, "kvPlaces": { "hbm": "显存", "host-cache": "主机内存缓存层", "host-stream": "主机内存常驻" } }