{
  "arch_id": "qwen3-next-mtp",
  "artifact_role": "published-derivative",
  "base_trunk": "Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
  "exactness_baseline": {},
  "forge_provenance": {
    "forge_inputs": {
      "mtp_source": "Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
      "trunk_source": "Qwen/Qwen3.8-27B@1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 plus PocketAiHub refusal-direction projection"
    },
    "forge_recipe": {
      "body_bits": 4,
      "body_dtype": "bf16",
      "body_group_size": 32,
      "body_mode": "affine",
      "forge_version": "2.7.1",
      "module_overrides": [
        {
          "bits": 8,
          "group_size": 64,
          "suffix": "embed_tokens"
        },
        {
          "bits": 8,
          "group_size": 64,
          "suffix": "lm_head"
        },
        {
          "bits": 8,
          "group_size": 64,
          "suffix": "linear_attn.out_proj"
        },
        {
          "bits": 8,
          "group_size": 64,
          "layers": [
            56,
            57,
            58,
            59,
            60,
            61,
            62,
            63
          ],
          "suffix": "mlp.gate_proj"
        },
        {
          "bits": 8,
          "group_size": 64,
          "layers": [
            56,
            57,
            58,
            59,
            60,
            61,
            62,
            63
          ],
          "suffix": "mlp.up_proj"
        },
        {
          "bits": 8,
          "group_size": 64,
          "layers": [
            56,
            57,
            58,
            59,
            60,
            61,
            62,
            63
          ],
          "suffix": "mlp.down_proj"
        }
      ],
      "mtp_policy": "keep_bf16",
      "recipe_origin": "Youssofal/Qwen3.8-27B-MTPLX-Optimized-Speed",
      "source_repo": "Qwen/Qwen3.8-27B",
      "source_revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0"
    },
    "forged_at": "2026-08-16T13:35:38+09:00",
    "forged_locally": true,
    "mtp_contract": {
      "base_hidden_variant": "post_norm",
      "concat_order": "embedding_hidden",
      "hidden_variant": "post_norm",
      "mtp_position_mode": "local",
      "mtp_quant_group_size": 64,
      "mtp_quant_mode": "affine"
    },
    "mtplx_version": "2.7.1",
    "published_to_hf": "PocketAiHub/Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed",
    "source_format": "bf16_native",
    "source_repo": "Qwen/Qwen3.8-27B",
    "source_sha": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
    "recommended_profile_note": "Turbo validated on Apple M5 Max with Apple automatic fan control; official Qwen3.8 sampler selected D3."
  },
  "mtp_contract": {
    "base_hidden_variant": "post_norm",
    "concat_order": "embedding_hidden",
    "hidden_variant": "post_norm",
    "mtp_position_mode": "local",
    "mtp_quant_group_size": 64,
    "mtp_quant_mode": "affine"
  },
  "mtp_depth_max": 3,
  "mtp_sidecar": "bf16",
  "mtplx_version": "2.7.1",
  "recommended_profile": "turbo",
  "sampler": {
    "temperature": 0.6,
    "top_k": 20,
    "top_p": 0.95
  },
  "speed_evidence": {
    "source": "evaluation/results.json",
    "profile": "performance-cold",
    "sampler": {
      "temperature": 1.0,
      "topP": 0.95,
      "topK": 20,
      "thinking": "disabled"
    },
    "depth": 3,
    "tok_s": [
      58.053625774312664
    ],
    "multiplier_vs_ar": 2.346919375073558,
    "rows": [
      {
        "mode": "AR",
        "depth": null,
        "decodeTokensPerSecond": 24.736097196561396,
        "endToEndTokensPerSecond": 24.282636058208418,
        "speedupVsAR": 1.0,
        "acceptanceByDepth": null,
        "verifyMillisecondsPerCall": null,
        "generatedTokens": 418,
        "qualityPassed": true
      },
      {
        "mode": "D1",
        "depth": 1,
        "decodeTokensPerSecond": 41.54993385858086,
        "endToEndTokensPerSecond": 40.52622110824334,
        "speedupVsAR": 1.6797287594890589,
        "acceptanceByDepth": [
          0.956
        ],
        "verifyMillisecondsPerCall": 45.116196647526635,
        "generatedTokens": 512,
        "qualityPassed": true
      },
      {
        "mode": "D2",
        "depth": 2,
        "decodeTokensPerSecond": 51.549283031569075,
        "endToEndTokensPerSecond": 49.70212018655927,
        "speedupVsAR": 2.0839699416581783,
        "acceptanceByDepth": [
          0.9790209790209791,
          0.916083916083916
        ],
        "verifyMillisecondsPerCall": 50.65876786468615,
        "generatedTokens": 437,
        "qualityPassed": true
      },
      {
        "mode": "D3",
        "depth": 3,
        "decodeTokensPerSecond": 58.053625774312664,
        "endToEndTokensPerSecond": 56.01466873186657,
        "speedupVsAR": 2.346919375073558,
        "acceptanceByDepth": [
          0.9629629629629629,
          0.8880597014925373,
          0.8059701492537313
        ],
        "verifyMillisecondsPerCall": 53.64032379764146,
        "generatedTokens": 512,
        "qualityPassed": true
      }
    ]
  },
  "verified_on": {
    "hardware": "macOS-26.4-arm64-arm-64bit",
    "machine_arch": "arm64",
    "macos": "26.4",
    "model": "Qwen3.8-27B-Abliterated-MTPLX-Optimized-Speed",
    "timestamp": "2026-08-16T13:35:38+09:00"
  },
  "recommended_draft_sampler": {
    "temperature": 1.0,
    "top_p": 0.95,
    "top_k": 20
  },
  "release_validation": {
    "status": "passed",
    "context4K": true,
    "capability": "12/12",
    "tools": "8/8 selection",
    "visionImage": true,
    "reasoningOnOff": true,
    "explicitRefusal": "0/100 harmful and 0/100 benign"
  }
}
