{
  "schema_version": 2,
  "published_at": "2026-07-25",
  "device": {
    "gpu": "Intel Arc Pro B70 (Battlemage G31)",
    "reported_memory_mib": 32656,
    "kernel_driver": "xe",
    "xpu_runtime": "Level Zero",
    "xpu_device_count": 1,
    "xpu_compute_units": 256
  },
  "methodology": {
    "long_context": {
      "context": 131072,
      "depth": 115404,
      "prompt_tokens": 2048,
      "generated_tokens": 512,
      "target_fill_percent": 90,
      "concurrency": 1,
      "measured_runs": 1,
      "full_workload_warmup_runs": 0,
      "pre_measurement_checks": [
        "user-only warmup",
        "system-and-probe warmup",
        "fixed coherence check",
        "generation latency probe"
      ],
      "cache_reuse_between_cells": false,
      "notes": "Each reported cell is one measured run after the listed checks. These are observations, not estimates of a population mean."
    },
    "throughput_units": "tokens per second",
    "engine_tuning": "vLLM and llama.cpp used their own tested scheduler and batch settings. The comparison covers complete engine and weight-format paths."
  },
  "quantization": [
    {
      "family": "AutoRound INT4",
      "weight_bits": 4,
      "symmetric": true,
      "group_size": 128,
      "packing_format": "auto_round:auto_gptq",
      "notes": "Selected layers and gates remain at higher precision where listed by each checkpoint's quantization metadata."
    },
    {
      "family": "Gemma 4 QAT W4A16",
      "weight_bits": 4,
      "activation_bits": 16,
      "symmetric": true,
      "group_size": 32,
      "packing_format": "compressed-tensors pack-quantized"
    },
    {
      "family": "GGUF Q4_K_S",
      "weight_bits": 4,
      "packing_format": "GGUF K-quant",
      "notes": "The published quality suite does not compare this format with BF16 or AutoRound."
    }
  ],
  "startup_memory": [
    {
      "label": "Qwen3.6 27B AutoRound",
      "model_load_gib": 17.75,
      "available_kv_gib": 9.04,
      "kv_tokens": 144079,
      "gpu_memory_utilization": 0.9
    },
    {
      "label": "Qwen3.6 35B-A3B AutoRound",
      "model_load_gib": 19.6,
      "available_kv_gib": 5.69,
      "kv_tokens": 289626,
      "gpu_memory_utilization": 0.9
    },
    {
      "label": "Ornith 35B AutoRound",
      "model_load_gib": 19.06,
      "available_kv_gib": 6.23,
      "kv_tokens": 317109,
      "gpu_memory_utilization": 0.9
    },
    {
      "label": "Gemma 4 12B AutoRound",
      "model_load_gib": 7.63,
      "available_kv_gib": 19.13,
      "kv_tokens": 700098,
      "gpu_memory_utilization": 0.9
    },
    {
      "label": "Gemma 4 12B BF16",
      "model_load_gib": 22.73,
      "available_kv_gib": 7.58,
      "kv_tokens": 205455,
      "gpu_memory_utilization": 0.98
    }
  ]
}