[
  {
    "artifact_id": "qwen3.5-0.8b__int8hu__aimodel",
    "model": "Qwen3.5-0.8B",
    "vendor": "Qwen",
    "params_b": 0.8,
    "quant": "int8hu",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.3969,
      "mmlu_ci": [
        0.3172,
        0.4825
      ],
      "mmlu_acc": 0.2653,
      "mmlu_n": 196,
      "mmlu_answered": 131,
      "mmlu_noans": 65,
      "math_completed": 0.5926,
      "math_ci": [
        0.4597,
        0.7132
      ],
      "math_acc": 0.32,
      "math_n": 100,
      "math_answered": 54,
      "math_noans": 46,
      "ifeval_mean4": 0.6022,
      "ifeval_ci": [
        0.5515,
        0.6518
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.5306,
      "ci": [
        0.4736,
        0.5846
      ]
    },
    "retention": {
      "baseline": "fp16-engine(+eager anchor)",
      "metric": "completed-only",
      "mmlu": 0.938,
      "mmlu_baseline": 0.4737,
      "mmlu_baseline_n": 19,
      "math": 0.938,
      "math_baseline": 0.6,
      "math_baseline_n": 15,
      "ifeval": 0.898,
      "ifeval_baseline": 0.6708,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "port-time token-exact gate; device tok/s from earlier verified runs (zoo notes)"
    },
    "iphone_tok_s": 72,
    "mac_tok_s": 210,
    "mem_mb": 443,
    "mem_measured": true
  },
  {
    "artifact_id": "qwen3.5-2b__int8hu__aimodel",
    "model": "Qwen3.5-2B",
    "vendor": "Qwen",
    "params_b": 2.0,
    "quant": "int8hu",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.6098,
      "mmlu_ci": [
        0.5215,
        0.6914
      ],
      "mmlu_acc": 0.3827,
      "mmlu_n": 196,
      "mmlu_answered": 123,
      "mmlu_noans": 73,
      "math_completed": 0.8305,
      "math_ci": [
        0.7154,
        0.9052
      ],
      "math_acc": 0.49,
      "math_n": 100,
      "math_answered": 59,
      "math_noans": 41,
      "ifeval_mean4": 0.7029,
      "ifeval_ci": [
        0.6583,
        0.7474
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.7144,
      "ci": [
        0.6688,
        0.7611
      ]
    },
    "retention": {
      "baseline": "fp16-engine",
      "metric": "completed-only",
      "mmlu": 0.978,
      "mmlu_baseline": 0.625,
      "mmlu_baseline_n": 16,
      "math": 0.933,
      "math_baseline": 0.8333,
      "math_baseline_n": 18,
      "ifeval": 0.897,
      "ifeval_baseline": 0.7833,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "port-time token-exact gate; device tok/s from earlier verified runs (zoo notes)"
    },
    "iphone_tok_s": 29,
    "mac_tok_s": 165,
    "mem_mb": 2900,
    "mem_measured": false
  },
  {
    "artifact_id": "lfm2.5-1.2b__int8hu__aimodel",
    "model": "LFM2.5-1.2B",
    "vendor": "LiquidAI",
    "params_b": 1.2,
    "quant": "int8hu",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.4294,
      "mmlu_ci": [
        0.3559,
        0.5062
      ],
      "mmlu_acc": 0.3571,
      "mmlu_n": 196,
      "mmlu_answered": 163,
      "mmlu_noans": 33,
      "math_completed": 0.8507,
      "math_ci": [
        0.7466,
        0.9169
      ],
      "math_acc": 0.57,
      "math_n": 100,
      "math_answered": 67,
      "math_noans": 33,
      "ifeval_mean4": 0.8765,
      "ifeval_ci": [
        0.8428,
        0.9081
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.7189,
      "ci": [
        0.6776,
        0.7558
      ]
    },
    "retention": {
      "baseline": "fp16-engine(+transformers anchor)",
      "metric": "completed-only",
      "mmlu": 0.903,
      "mmlu_baseline": 0.4615,
      "mmlu_baseline_n": 26,
      "math": 0.955,
      "math_baseline": 0.9048,
      "math_baseline_n": 21,
      "ifeval": 1.032,
      "ifeval_baseline": 0.849,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "PB 2026-07-13 nat 24/24 oracle 24/24, iOS 24A5355q"
    },
    "iphone_tok_s": 45.5,
    "mac_tok_s": 277,
    "mem_mb": 1600,
    "mem_measured": false
  },
  {
    "artifact_id": "granite-4.0-h-1b__int8hu__aimodel",
    "model": "Granite-4.0-H-1B",
    "vendor": "IBM",
    "params_b": 1.0,
    "quant": "int8hu",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.2775,
      "mmlu_ci": [
        0.2188,
        0.3449
      ],
      "mmlu_acc": 0.2704,
      "mmlu_n": 196,
      "mmlu_answered": 191,
      "mmlu_noans": 5,
      "math_completed": 0.5667,
      "math_ci": [
        0.4636,
        0.6642
      ],
      "math_acc": 0.51,
      "math_n": 100,
      "math_answered": 90,
      "math_noans": 10,
      "ifeval_mean4": 0.7894,
      "ifeval_ci": [
        0.7485,
        0.8311
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.5445,
      "ci": [
        0.5002,
        0.5862
      ]
    },
    "retention": {
      "baseline": "fp16-engine",
      "metric": "completed-only",
      "mmlu": 0.963,
      "mmlu_baseline": 0.3462,
      "mmlu_baseline_n": 26,
      "math": 1.111,
      "math_baseline": 0.6,
      "math_baseline_n": 25,
      "ifeval": 0.936,
      "ifeval_baseline": 0.8438,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "PB 2026-07-13 nat 24/24 oracle 24/24, iOS 24A5355q"
    },
    "iphone_tok_s": 31,
    "mac_tok_s": 147,
    "mem_mb": 1700,
    "mem_measured": false
  },
  {
    "artifact_id": "qwen3.5-4b__int8hu__aimodel",
    "model": "Qwen3.5-4B",
    "vendor": "Qwen",
    "params_b": 4.0,
    "quant": "int8hu",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 4096,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.6889,
      "mmlu_ci": [
        0.5433,
        0.8047
      ],
      "mmlu_acc": 0.1582,
      "mmlu_n": 196,
      "mmlu_answered": 45,
      "mmlu_noans": 151,
      "math_completed": 0.5741,
      "math_ci": [
        0.4416,
        0.6967
      ],
      "math_acc": 0.31,
      "math_n": 100,
      "math_answered": 54,
      "math_noans": 46,
      "ifeval_mean4": 0.315,
      "ifeval_ci": [
        0.2711,
        0.3617
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.526,
      "ci": [
        0.461,
        0.5923
      ]
    },
    "retention": {
      "baseline": "fp16-engine",
      "metric": "completed-only",
      "mmlu": 0.778,
      "mmlu_baseline": 1.0,
      "mmlu_baseline_n": 8,
      "math": 0.938,
      "math_baseline": 0.5333,
      "math_baseline_n": 15,
      "ifeval": 1.219,
      "ifeval_baseline": 0.2583,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "PB 2026-07-13 24/24 (Mac-engine-anchored), iOS 24A5355q"
    },
    "note": "reasoning-heavy: cap4096",
    "iphone_tok_s": 10.8,
    "mac_tok_s": 45,
    "mem_mb": 4200,
    "mem_measured": false
  },
  {
    "artifact_id": "youtu-2b__int8__aimodel",
    "model": "Youtu-LLM-2B",
    "vendor": "Tencent",
    "params_b": 2.0,
    "quant": "int8",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 4096,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.6212,
      "mmlu_ci": [
        0.5361,
        0.6994
      ],
      "mmlu_acc": 0.4184,
      "mmlu_n": 196,
      "mmlu_answered": 132,
      "mmlu_noans": 64,
      "math_completed": 0.8983,
      "math_ci": [
        0.7954,
        0.9526
      ],
      "math_acc": 0.53,
      "math_n": 100,
      "math_answered": 59,
      "math_noans": 41,
      "ifeval_mean4": 0.6274,
      "ifeval_ci": [
        0.5728,
        0.6813
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.7156,
      "ci": [
        0.6739,
        0.7552
      ]
    },
    "retention": {
      "baseline": "eager-fp16 anchor (24 items, cap4096)",
      "metric": "completed-only",
      "mmlu": 1.543,
      "mmlu_baseline": 0.5833,
      "mmlu_baseline_n": 12,
      "math": 1.5,
      "math_baseline": 0.6667,
      "math_baseline_n": 12
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "PB 2026-07-13 16/16 (fp32-HF-anchored), iOS 24A5355q"
    },
    "note": "reasoning: cap4096. ② = eager small-n anchor (eager decodes 3.5 tok/s; 16 completed shared items): no quantization loss detected at that n",
    "iphone_tok_s": 23.3,
    "mac_tok_s": null,
    "mem_mb": 2100,
    "mem_measured": false
  },
  {
    "artifact_id": "gemma-4-e2b__int4__aimodel",
    "model": "Gemma 4 E2B",
    "vendor": "Google",
    "params_b": 5.4,
    "quant": "int4",
    "format": "aimodel",
    "native_runtime": "coreai",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.5294,
      "mmlu_ci": [
        0.4459,
        0.6113
      ],
      "mmlu_acc": 0.3673,
      "mmlu_n": 196,
      "mmlu_answered": 136,
      "mmlu_noans": 60,
      "math_completed": 0.8378,
      "math_ci": [
        0.6886,
        0.9235
      ],
      "math_acc": 0.31,
      "math_n": 100,
      "math_answered": 37,
      "math_noans": 63,
      "ifeval_mean4": 0.7618,
      "ifeval_ci": [
        0.7199,
        0.8042
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.7097,
      "ci": [
        0.6593,
        0.7587
      ]
    },
    "retention": {
      "baseline": "fp16-engine (official-QAT unquantized)",
      "metric": "completed-only",
      "mmlu": 1.05,
      "mmlu_baseline": 0.5238,
      "mmlu_baseline_n": 21,
      "math": 1.0,
      "math_baseline": 0.6667,
      "math_baseline_n": 6,
      "ifeval": 0.9,
      "ifeval_baseline": 0.8469,
      "ifeval_baseline_n": 40
    },
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": "PB 2026-07-13 8/8 chat-format greedy, PLE tables bound, iOS 24A5355q"
    },
    "note": "official Google QAT int4 (q4_0 grid): retention = QAT int4 vs its bf16 checkpoint; PLE tables int8 in BOTH columns (common-mode). E2B = 2B-effective (5.4B raw, per-layer embeds off-loaded)",
    "iphone_tok_s": 21.7,
    "mac_tok_s": 83,
    "mem_mb": 4450,
    "mem_measured": true
  },
  {
    "artifact_id": "apple-fm__system__system",
    "model": "Apple Foundation Model",
    "vendor": "Apple",
    "params_b": null,
    "quant": "system",
    "format": "system",
    "native_runtime": "foundation-models",
    "quality": {
      "battery_version": "v0",
      "cap_tokens": 1024,
      "shots": 0,
      "source": {
        "mmlu": "full596",
        "math": "full596",
        "ifeval": "full596"
      },
      "refusal_rate": 0.0,
      "refused_n": 0,
      "mmlu_completed": 0.648,
      "mmlu_ci": [
        0.561,
        0.7262
      ],
      "mmlu_acc": 0.4133,
      "mmlu_n": 196,
      "mmlu_answered": 125,
      "mmlu_noans": 71,
      "math_completed": 0.8148,
      "math_ci": [
        0.6916,
        0.8962
      ],
      "math_acc": 0.44,
      "math_n": 100,
      "math_answered": 54,
      "math_noans": 46,
      "ifeval_mean4": 0.8207,
      "ifeval_ci": [
        0.7799,
        0.8618
      ],
      "ifeval_n": 300
    },
    "composite": {
      "value": 0.7612,
      "ci": [
        0.7136,
        0.8067
      ]
    },
    "retention": null,
    "provenance": {
      "quality_run": "full596, Mac engine b2 toolchain (macOS 27 26A5378j), 2026-07-12/13",
      "device_gate": null
    },
    "iphone_tok_s": null,
    "mac_tok_s": null,
    "mem_mb": null,
    "mem_measured": false
  }
]