{
  "hardware": {
    "kind": "apple",
    "name": "M4 Max",
    "total_bytes": 137438953472,
    "usable_bytes": 103079215104,
    "bandwidth_gbps": 546.0,
    "cores": 16,
    "devices": 1,
    "note": "",
    "total_gb": 128.0,
    "usable_gb": 96.0
  },
  "context": 32768,
  "concurrency": 8,
  "feasible": [
    {
      "id": "qwen3-32b",
      "quant": "q4",
      "total_gb": 84.1,
      "headroom_gb": 11.9,
      "tokens_per_sec_estimate": 15,
      "estimate_basis": "memory-bandwidth roofline, not measured",
      "license": "Apache-2.0",
      "license_clean_commercial": true,
      "architecture_verified": true
    },
    {
      "id": "qwen3-30b-a3b",
      "quant": "q8",
      "total_gb": 56.2,
      "headroom_gb": 39.8,
      "tokens_per_sec_estimate": 60,
      "estimate_basis": "memory-bandwidth roofline, not measured",
      "license": "Apache-2.0",
      "license_clean_commercial": true,
      "architecture_verified": true
    },
    {
      "id": "mistral-small-24b",
      "quant": "q8",
      "total_gb": 65.2,
      "headroom_gb": 30.8,
      "tokens_per_sec_estimate": 14,
      "estimate_basis": "memory-bandwidth roofline, not measured",
      "license": "Apache-2.0",
      "license_clean_commercial": true,
      "architecture_verified": true
    },
    {
      "id": "phi-4-14b",
      "quant": "q8",
      "total_gb": 66.3,
      "headroom_gb": 29.7,
      "tokens_per_sec_estimate": 18,
      "estimate_basis": "memory-bandwidth roofline, not measured",
      "license": "MIT",
      "license_clean_commercial": true,
      "architecture_verified": true
    },
    {
      "id": "llama-3.1-8b",
      "quant": "q8",
      "total_gb": 41.6,
      "headroom_gb": 54.4,
      "tokens_per_sec_estimate": 32,
      "estimate_basis": "memory-bandwidth roofline, not measured",
      "license": "Llama 3.1",
      "license_clean_commercial": false,
      "architecture_verified": true
    }
  ],
  "rejected": [
    {
      "id": "gemma-3-27b",
      "reason": "weights fit (14 GB) but KV at 32,768 ctx x8 (124 GB) puts it 45 GB over the 96 GB available"
    },
    {
      "id": "llama-3.3-70b",
      "reason": "weights fit (37 GB) but KV at 32,768 ctx x8 (80 GB) puts it 25 GB over the 96 GB available"
    },
    {
      "id": "qwen3-235b-a22b",
      "reason": "weights alone need 123 GB at q4 \u2014 MoE sparsity (22B of 235B active) cuts compute, not memory"
    },
    {
      "id": "deepseek-v3",
      "reason": "weights alone need 352 GB at q4 \u2014 MoE sparsity (37B of 671B active) cuts compute, not memory"
    },
    {
      "id": "glm-5.2",
      "reason": "weights alone need 186 GB at q4 \u2014 MoE sparsity (32B of 355B active) cuts compute, not memory"
    },
    {
      "id": "kimi-k2.7-code",
      "reason": "weights alone need 524 GB at q4 \u2014 MoE sparsity (32B of 1000B active) cuts compute, not memory"
    },
    {
      "id": "kimi-k3",
      "reason": "weights alone need 1,467 GB at q4 \u2014 MoE sparsity (50B of 2800B active) cuts compute, not memory"
    },
    {
      "id": "deepseek-v4-pro",
      "reason": "weights alone need 838 GB at q4 \u2014 MoE sparsity (45B of 1600B active) cuts compute, not memory"
    },
    {
      "id": "minimax-m3",
      "reason": "weights alone need 239 GB at q4 \u2014 MoE sparsity (46B of 456B active) cuts compute, not memory"
    }
  ],
  "recommended_runtime": {
    "name": "mlx",
    "why": "Apple silicon: the CUDA engines cannot run here at all. MLX has the better batching story of the two Metal options."
  }
}
