{ "source": "https://tensorlight.dev/models", "generated": "2026-11-15", "constants": { "ETA_BW": 0.65, "ETA_C": 0.5, "LAYER_S": 0.00006, "LAYER_SYNC_S": 0.00004, "STAGE_HOP_S": 0.00005 }, "devices": [ { "id": "h100", "name": "H100", "segment": "datacenter", "memGB": 80, "usable": 0.9, "bwTBs": 3.35, "bf16": 989, "fp8": 1979, "fp4": null, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "h200", "name": "H200", "segment": "datacenter", "memGB": 141, "usable": 0.9, "bwTBs": 4.8, "bf16": 989, "fp8": 1979, "fp4": null, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "b200", "name": "B200", "segment": "datacenter", "memGB": 180, "usable": 0.9, "bwTBs": 8, "bf16": 2250, "fp8": 4500, "fp4": 9000, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "b300", "name": "B300", "segment": "datacenter", "memGB": 270, "usable": 0.9, "bwTBs": 8, "bf16": 2250, "fp8": 4500, "fp4": 13500, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "mi300x", "name": "MI300X", "segment": "datacenter", "memGB": 192, "usable": 0.9, "bwTBs": 5.3, "bf16": 1307, "fp8": 2615, "fp4": null, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "mi355x", "name": "MI355X", "segment": "datacenter", "memGB": 288, "usable": 0.9, "bwTBs": 8, "bf16": 2516, "fp8": 5033, "fp4": 10066, "scaleUp": 8, "maxGroup": 128, "grouping": "server" }, { "id": "rtx-pro-6000", "name": "RTX PRO 6000 Blackwell", "segment": "personal", "memGB": 96, "usable": 0.9, "bwTBs": 1.792, "bf16": 504, "fp8": 1008, "fp4": 2000, "scaleUp": 4, "maxGroup": 4, "grouping": "any" }, { "id": "rtx-5090", "name": "GeForce RTX 5090", "segment": "personal", "memGB": 32, "usable": 0.9, "bwTBs": 1.792, "bf16": 209.5, "fp8": 419, "fp4": 1676, "scaleUp": 4, "maxGroup": 4, "grouping": "any" }, { "id": "rtx-4090", "name": "GeForce RTX 4090 (used)", "segment": "personal", "memGB": 24, "usable": 0.9, "bwTBs": 1.008, "bf16": 165.2, "fp8": 330.3, "fp4": null, "scaleUp": 4, "maxGroup": 4, "grouping": "any" }, { "id": "dgx-spark", "name": "DGX Spark", "segment": "personal", "memGB": 128, "usable": 0.85, "bwTBs": 0.273, "bf16": 125, "fp8": 250, "fp4": 500, "scaleUp": 2, "maxGroup": 2, "grouping": "any" }, { "id": "mac-studio-m5-ultra", "name": "Mac Studio M5 Ultra (512 GB)", "segment": "personal", "memGB": 512, "usable": 0.75, "bwTBs": 1.2, "bf16": 140, "fp8": null, "fp4": null, "scaleUp": 1, "maxGroup": 1, "grouping": "any" }, { "id": "mac-mini-m5-pro", "name": "Mac mini M5 Pro (64 GB)", "segment": "personal", "memGB": 64, "usable": 0.75, "bwTBs": 0.307, "bf16": 35, "fp8": null, "fp4": null, "scaleUp": 1, "maxGroup": 1, "grouping": "any" }, { "id": "strix-halo", "name": "Ryzen AI Max+ 395 (128 GB)", "segment": "personal", "memGB": 128, "usable": 0.85, "bwTBs": 0.256, "bf16": 59.4, "fp8": null, "fp4": null, "scaleUp": 1, "maxGroup": 1, "grouping": "any" } ], "known": { "moonshotai/kimi-k3": { "activeB": 104, "kvPerToken": 13824, "kvWindowed": 0, "window": 0, "attention": "24 MLA layers + 69 linear (KDA) layers", "exact": true }, "moonshotai/kimi-k3.1-pro": { "activeB": 104, "kvPerToken": 16896, "kvWindowed": 0, "window": 0, "attention": "24 sparse MLA (DSA) layers + 69 linear (KDA) layers", "exact": true }, "qwen/qwen3.8-2.4t-a95b": { "activeB": 95, "kvPerToken": 47104, "kvWindowed": 0, "window": 0, "attention": "23 attention + 69 linear (Gated DeltaNet) layers", "exact": true }, "deepseek-ai/deepseek-v4-pro": { "activeB": 49, "kvPerToken": 4924, "kvWindowed": 31232, "window": 128, "attention": "compressed attention (30 layers at 1/4, 31 at 1/128) + 128-token window", "exact": false }, "xiaomimimo/mimo-v2.6-pro-rl": { "activeB": 42, "kvPerToken": 25600, "kvWindowed": 153600, "window": 128, "attention": "10 full + 60 sliding-window (128) layers", "exact": true }, "moonshotai/kimi-k2.6": { "activeB": 32, "kvPerToken": 35136, "kvWindowed": 0, "window": 0, "attention": "MLA on all 61 layers", "exact": true }, "mistralai/mistral-large-4-1t-a52b": { "activeB": 52, "kvPerToken": 41920, "kvWindowed": 0, "window": 0, "attention": "MLA + indexer; 63 of 72 layers keep their own KV", "exact": true }, "thinkingmachines/inkling": { "activeB": 41, "kvPerToken": 22528, "kvWindowed": 225280, "window": 512, "attention": "11 global + 55 local (512-token) layers", "exact": true }, "zai-org/glm-5.3": { "activeB": 40, "kvPerToken": 47616, "kvWindowed": 0, "window": 0, "attention": "MLA on all 78 layers; sparse-attention indexer on 21", "exact": true }, "zai-org/glm-5": { "activeB": 40, "kvPerToken": 54912, "kvWindowed": 0, "window": 0, "attention": "MLA + sparse-attention indexer on all 78 layers", "exact": true }, "deepseek-ai/deepseek-v3.2": { "activeB": 37, "kvPerToken": 42944, "kvWindowed": 0, "window": 0, "attention": "MLA + sparse-attention indexer on all 61 layers", "exact": true }, "deepseek-ai/deepseek-v4.1-flash-1017": { "activeB": 16, "kvPerToken": 1600, "kvWindowed": 20480, "window": 128, "attention": "4 shared compressed KV caches (3 at 1/2, 1 full) + 128-token window", "exact": false }, "nvidia/nvidia-nemotron-3-ultra-550b-a55b-bf16": { "activeB": 55, "kvPerToken": 6144, "kvWindowed": 0, "window": 0, "attention": "12 attention layers among 108 Mamba-2 / MoE layers", "exact": true }, "reflection/beam-1.0-501b-a23b": { "activeB": 23, "kvPerToken": 18432, "kvWindowed": 88064, "window": 512, "attention": "9 full + 43 sliding-window (512) layers", "exact": true }, "minimaxai/minimax-m3": { "activeB": 23, "kvPerToken": 61440, "kvWindowed": 0, "window": 0, "attention": "GQA on all 60 layers with block-sparse attention", "exact": false }, "minimaxai/minimax-m3.1-flash": { "activeB": 23, "kvPerToken": 61440, "kvWindowed": 0, "window": 0, "attention": "GQA on all 60 layers with block-sparse attention", "exact": false }, "qwen/qwen3.5-397b-a17b": { "activeB": 17, "kvPerToken": 15360, "kvWindowed": 0, "window": 0, "attention": "15 attention + 45 linear (Gated DeltaNet) layers", "exact": true }, "zai-org/glm-5.3-flash": { "activeB": 18, "kvPerToken": 7040, "kvWindowed": 0, "window": 0, "attention": "11 sparse MLA layers + 34 linear (KDA) layers", "exact": true }, "xiaomimimo/mimo-v2.6-flash-rl": { "activeB": 15, "kvPerToken": 11520, "kvWindowed": 99840, "window": 128, "attention": "9 full + 39 sliding-window (128) layers", "exact": true }, "tencent/hy3": { "activeB": 21, "kvPerToken": 163840, "kvWindowed": 0, "window": 0, "attention": "GQA on all 80 layers (8 KV heads)", "exact": true }, "deepseek-ai/deepseek-v4-flash": { "activeB": 13, "kvPerToken": 3440, "kvWindowed": 22016, "window": 128, "attention": "compressed attention (21 layers at 1/4, 20 at 1/128) + 128-token window", "exact": false }, "stepfun-ai/step-3.7-flash": { "activeB": 11, "kvPerToken": 24576, "kvWindowed": 67584, "window": 512, "attention": "12 full + 33 sliding-window (512) layers", "exact": true }, "qwen/qwen3.8-flash-next": { "activeB": 6, "kvPerToken": 12288, "kvWindowed": 0, "window": 0, "attention": "12 attention + 36 linear (Gated DeltaNet) layers", "exact": false }, "inclusionai/ling-3.0-flash": { "activeB": 5.1, "kvPerToken": 4032, "kvWindowed": 0, "window": 0, "attention": "about 7 MLA layers + 35 linear (KDA) layers", "exact": false }, "qwen/qwen3.5-122b-a10b": { "activeB": 10, "kvPerToken": 12288, "kvWindowed": 0, "window": 0, "attention": "12 attention + 36 linear (Gated DeltaNet) layers", "exact": true }, "nvidia/nvidia-nemotron-3-super-120b-a12b-fp8": { "activeB": 12, "kvPerToken": 4096, "kvWindowed": 0, "window": 0, "attention": "8 attention layers among 88 Mamba-2 / MoE layers", "exact": true }, "openai/gpt-oss-120b": { "activeB": 5.1, "kvPerToken": 18432, "kvWindowed": 18432, "window": 128, "attention": "18 full + 18 sliding-window (128) layers", "exact": true }, "zai-org/glm-4.5-air": { "activeB": 12, "kvPerToken": 94208, "kvWindowed": 0, "window": 0, "attention": "GQA on all 46 layers (8 KV heads)", "exact": true }, "bytedance-seed/seed-oss-36b-instruct": { "activeB": 36, "kvPerToken": 131072, "kvWindowed": 0, "window": 0, "attention": "GQA on all 64 layers (8 KV heads)", "exact": true }, "qwen/qwen3.6-35b-a3b": { "activeB": 3, "kvPerToken": 10240, "kvWindowed": 0, "window": 0, "attention": "10 attention + 30 linear (Gated DeltaNet) layers", "exact": true }, "google/gemma-4-31b-it": { "activeB": 31, "kvPerToken": 20480, "kvWindowed": 409600, "window": 1024, "attention": "10 global + 50 sliding-window (1024) layers", "exact": true }, "zai-org/glm-4.7-flash": { "activeB": 3, "kvPerToken": 27072, "kvWindowed": 0, "window": 0, "attention": "MLA on all 47 layers", "exact": true }, "nvidia/nvidia-nemotron-3-nano-30b-a3b-bf16": { "activeB": 3.5, "kvPerToken": 3072, "kvWindowed": 0, "window": 0, "attention": "6 attention layers among 52 Mamba-2 / MoE layers", "exact": true }, "meta-models/muse-glimmer-30b": { "activeB": 29.6, "kvPerToken": 6656, "kvWindowed": 19968, "window": 2048, "attention": "13 full + 39 sliding-window (2048) layers", "exact": true }, "qwen/qwen3.8-27b": { "activeB": 27, "kvPerToken": 32768, "kvWindowed": 0, "window": 0, "attention": "16 attention + 48 linear (Gated DeltaNet) layers", "exact": true }, "google/gemma-4-26b-a4b-it": { "activeB": 3.8, "kvPerToken": 5120, "kvWindowed": 102400, "window": 1024, "attention": "5 global + 25 sliding-window (1024) layers", "exact": true }, "openai/gpt-oss-20b": { "activeB": 3.6, "kvPerToken": 12288, "kvWindowed": 12288, "window": 128, "attention": "12 full + 12 sliding-window (128) layers", "exact": true }, "qwen/qwen3.5-9b": { "activeB": 9, "kvPerToken": 16384, "kvWindowed": 0, "window": 0, "attention": "8 attention + 24 linear (Gated DeltaNet) layers", "exact": true }, "qwen/qwen3.5-4b": { "activeB": 4, "kvPerToken": 16384, "kvWindowed": 0, "window": 0, "attention": "8 attention + 24 linear (Gated DeltaNet) layers", "exact": true } } }