{
  "status": "complete",
  "model": "Qwen/Qwen3.5-9B",
  "device": "single P150",
  "checkpoint": "New standalone mixed-BFP4/BFP8 TT-native checkpoint including all 15 original MTP tensors",
  "runtime_image": "sha256:fe06066653f6c1dc3d68cf7c2ee6e31a7724aebec82f88ff7bf792b78a0b1b21",
  "manifest_sha256": "34f774cf110e3ff94726171b52f6d462cb9b36651295b941efa85098ddb96ca2",
  "target_tensors_byte_identical_to_previous_release": 475,
  "mtp_source_tensors": 15,
  "mtp_source_bytes": 486582864,
  "native_reload": [
    {
      "mode": "off",
      "file": "equivalence.json",
      "exact_logits_equal": true,
      "tokens_compared": 1415
    },
    {
      "mode": "on",
      "file": "equivalence-mtp.json",
      "exact_logits_equal": true,
      "tokens_compared": 1415
    }
  ],
  "api": {
    "behavior_passes": 11,
    "long_context_passes": 1,
    "request_isolation_passed": true
  },
  "quality": {
    "gsm8k": {
      "accuracyPct": 96.04,
      "avgLatencyMs": 7899,
      "correct": 97,
      "dataset": "gsm8k",
      "declaredModel": "",
      "effectiveModel": "",
      "errors": 0,
      "quantFormat": "ttnn-native",
      "quantization": "mixed-bfp4-bfp8",
      "questions": 101,
      "scored": 101,
      "scoring": "exact_match",
      "shardIndex": 1
    },
    "arc_challenge": {
      "accuracyPct": 94.25,
      "avgLatencyMs": 218,
      "correct": 754,
      "dataset": "arc-challenge",
      "declaredModel": "",
      "effectiveModel": "",
      "errors": 0,
      "quantFormat": "ttnn-native",
      "quantization": "mixed-bfp4-bfp8",
      "questions": 800,
      "scored": 800,
      "scoring": "exact_match",
      "shardIndex": 1
    }
  },
  "speed": {
    "fixed_prompt": {
      "off_tok_s": 36.571853797115395,
      "on_tok_s": 50.564418615171256,
      "speedup_percent": 38.26047455969959,
      "runs_per_mode": 5
    },
    "localmaxxing": {
      "off_tok_s": 36.5,
      "on_tok_s": 55.1,
      "speedup_percent": 50.95890410958906,
      "runs_per_mode": 3
    },
    "draft_acceptance": {
      "accepted_per_run": 47,
      "proposed_per_run": 80,
      "percent": 58.75
    }
  },
  "caveats": [
    "Greedy decoding, concurrency 1, thinking disabled.",
    "MTP-on and MTP-off completions differ; reload equivalence is not cross-mode output equivalence.",
    "Quality scores are complete declared LocalMaxxing shards, not full-dataset official scores.",
    "Previous no-MTP quality scores also use an older runtime; comparison does not isolate MTP alone.",
    "Nothing submitted publicly."
  ],
  "serving_restored": {
    "base_url": "http://127.0.0.1:8000",
    "health_status": 200,
    "container": "qwen35-tt-bfp4-hf",
    "image": "sha256:3c76d5443033f2245a3fc2edebdeb6c520ed6f392f3a081991fc9e1a27e1b95d",
    "note": "Original endpoint restored unchanged; fixed new quant is packaged separately and defaults to MTP on."
  }
}