{
  "models": {
    "inkling-small": {
      "architecture": "inkling",
      "blocks": 42,
      "draft_blocks": 0,
      "experts": 256,
      "active_experts": 6,
      "cpu_expert_layers_setting": 42,
      "file_bytes": 119554379840,
      "category_stored_bytes": {
        "output head and output norms": 675457024,
        "input embedding": 874856448,
        "routing, norms and other weights": 169828520,
        "attention and recurrent state weights": 1974511616,
        "dense feed-forward": 427819008,
        "routed experts": 113279762432,
        "shared experts and shared gate": 2139095040
      },
      "expert_RAM_bytes_per_token": 2654994432,
      "GPU_active_expert_bytes_per_token": 0,
      "input_embedding_row_bytes": 4352,
      "pair_remote_expert_bytes_per_token": 0,
      "pair_local_expert_bytes_per_token": 2654994432
    },
    "qwen397": {
      "architecture": "qwen35moe",
      "blocks": 61,
      "draft_blocks": 1,
      "experts": 512,
      "active_experts": 10,
      "cpu_expert_layers_setting": 57,
      "file_bytes": 149840422542,
      "category_stored_bytes": {
        "output head and output norms": 834371584,
        "input embedding": 834355200,
        "attention and recurrent state weights": 5750041600,
        "routed experts": 138814685184,
        "shared experts and shared gate": 623345664,
        "routing, norms and other weights": 504299520,
        "draft weights": 2468308992
      },
      "expert_RAM_bytes_per_token": 2563112960,
      "GPU_active_expert_bytes_per_token": 148111360,
      "input_embedding_row_bytes": 3360,
      "pair_remote_expert_bytes_per_token": 0,
      "pair_local_expert_bytes_per_token": 2563112960
    },
    "deepseek": {
      "architecture": "deepseek4",
      "blocks": 43,
      "draft_blocks": 0,
      "experts": 256,
      "active_experts": 6,
      "cpu_expert_layers_setting": 43,
      "file_bytes": 161869615520,
      "category_stored_bytes": {
        "output head and output norms": 1059340308,
        "input embedding": 1059061760,
        "attention and recurrent state weights": 10243999524,
        "routed experts": 147169738752,
        "shared experts and shared gate": 2164260864,
        "routing, norms and other weights": 167868964
      },
      "expert_RAM_bytes_per_token": 3449290752,
      "GPU_active_expert_bytes_per_token": 0,
      "input_embedding_row_bytes": 8192,
      "pair_remote_expert_bytes_per_token": 0,
      "pair_local_expert_bytes_per_token": 3449290752
    },
    "inkling-975b": {
      "architecture": "inkling",
      "blocks": 66,
      "draft_blocks": 0,
      "experts": 256,
      "active_experts": 6,
      "cpu_expert_layers_setting": 66,
      "file_bytes": 270163818071,
      "category_stored_bytes": {
        "output head and output norms": 694763520,
        "input embedding": 694738944,
        "routing, norms and other weights": 407511304,
        "attention and recurrent state weights": 6091720704,
        "dense feed-forward": 662962176,
        "routed experts": 256276168704,
        "shared experts and shared gate": 5322866688
      },
      "expert_RAM_bytes_per_token": 6006472704,
      "GPU_active_expert_bytes_per_token": 0,
      "input_embedding_row_bytes": 3456,
      "pair_remote_expert_bytes_per_token": 2358706176,
      "pair_local_expert_bytes_per_token": 3647766528
    }
  },
  "anchors": [
    {
      "model": "inkling-small",
      "record": "runs/inkling-small-prose-r1.json",
      "repetition": 1,
      "rate": 11.772442551066042,
      "generated_tokens": 202,
      "prompt_tokens": 39,
      "visible_words": 174,
      "finish_reason": "stop",
      "selected": true,
      "effective_expert_GB_s": 31.25576942412022
    },
    {
      "model": "inkling-small",
      "record": "runs/inkling-small-prose-r2.json",
      "repetition": 2,
      "rate": 11.692801143730472,
      "generated_tokens": 202,
      "prompt_tokens": 39,
      "visible_words": 174,
      "finish_reason": "stop",
      "selected": true,
      "effective_expert_GB_s": 31.044321931087634
    },
    {
      "model": "qwen397",
      "record": "runs/qwen397-prose-r1.json",
      "repetition": 1,
      "rate": 12.757664964581535,
      "generated_tokens": 184,
      "prompt_tokens": 40,
      "visible_words": 161,
      "finish_reason": "stop",
      "selected": true,
      "effective_expert_GB_s": 32.69933641005687
    },
    {
      "model": "qwen397",
      "record": "runs/qwen397-prose-r2.json",
      "repetition": 2,
      "rate": 12.844951522175622,
      "generated_tokens": 184,
      "prompt_tokens": 40,
      "visible_words": 161,
      "finish_reason": "stop",
      "selected": true,
      "effective_expert_GB_s": 32.923061717060065
    },
    {
      "model": "deepseek",
      "record": "runs/deepseek-prose-r1.json",
      "repetition": 1,
      "rate": 2.0704385586113965,
      "generated_tokens": 209,
      "prompt_tokens": 30,
      "visible_words": 166,
      "finish_reason": "stop",
      "selected": false,
      "effective_expert_GB_s": 7.1415445728025
    },
    {
      "model": "deepseek",
      "record": "runs/deepseek-prose-r2.json",
      "repetition": 2,
      "rate": 9.933466685766806,
      "generated_tokens": 209,
      "prompt_tokens": 30,
      "visible_words": 166,
      "finish_reason": "stop",
      "selected": true,
      "effective_expert_GB_s": 34.26341477451553
    }
  ],
  "effective_expert_GB_s_range": [
    31.044321931087634,
    34.26341477451553
  ],
  "inkling975_prediction_tokens_s": [
    5.168477983827967,
    5.704415297134019
  ],
  "units": "GB is decimal: 1000000000 bytes. All bandwidth and prediction values are arithmetic."
}