{
  "model_name": "qwen3.5:4b",
  "quantization_level": "Q4_K_M",
  "physical_model_size_gb": 3.16,
  "estimated_fp16_size_gb": 8.6,
  "memory_footprint_reduction_x": 2.72,
  "accuracy_count": 4,
  "num_tasks": 6,
  "accuracy_pct": 66.67,
  "mean_decode_tokens_per_sec": 78.7,
  "outcomes": [
    {
      "task_id": "QF-01-INT4-VRAM-BUDGET",
      "topic": "4-Bit Weight + Group-Scale Storage",
      "expected_value": 9,
      "predicted_value": 9,
      "exact_match": true,
      "eval_tokens": 159,
      "tokens_per_sec": 78.4
    },
    {
      "task_id": "QF-02-KV-CACHE-GQA-RATIO",
      "topic": "Grouped-Query Attention KV Compression",
      "expected_value": 64,
      "predicted_value": 25,
      "exact_match": false,
      "eval_tokens": 128,
      "tokens_per_sec": 78.4
    },
    {
      "task_id": "QF-03-ROOFLINE-RIDGE",
      "topic": "Hardware Roofline Ridge Intensity",
      "expected_value": 200,
      "predicted_value": 24,
      "exact_match": false,
      "eval_tokens": 117,
      "tokens_per_sec": 78.2
    },
    {
      "task_id": "QF-04-FP8-E4M3-MAX",
      "topic": "OCP FP8 Dynamic Range Spec",
      "expected_value": 448,
      "predicted_value": 448,
      "exact_match": true,
      "eval_tokens": 64,
      "tokens_per_sec": 78.5
    },
    {
      "task_id": "QF-05-TENSOR-PARALLEL-SHARD",
      "topic": "8-Way Tensor Parallelism Memory",
      "expected_value": 6,
      "predicted_value": 6,
      "exact_match": true,
      "eval_tokens": 48,
      "tokens_per_sec": 79.3
    },
    {
      "task_id": "QF-06-HADAMARD-NORM",
      "topic": "Orthonormal Walsh-Hadamard L2 Preservation",
      "expected_value": 361,
      "predicted_value": 361,
      "exact_match": true,
      "eval_tokens": 88,
      "tokens_per_sec": 79.2
    }
  ]
}
