{
  "format": "neural-download-model-family-v1",
  "id": "qwen-35b",
  "primary_packet_id": "qwen36-35b-quark-int8-b70-result",
  "name": "Qwen 35B",
  "display_name": "Qwen3.6 35B A3B",
  "summary": "Alibaba's Qwen3.6 35B mixture-of-experts (about 3B active per word), built for long documents. Measured on four cards under the lab's strictest quality gate, on a single card experimentally, and serving 64 people at once from one card.",
  "updated_at": "2026-08-24",
  "featured_results": [
    {
      "role": "hero",
      "label": "Quark W8A8 INT8 · TP4 strict reference",
      "measurement_id": "qwen35-quark-tp4-strict-current",
      "metric": "decode_tok_s",
      "sample_index": 0,
      "quality_label": "Current strict deep-gate reference"
    },
    {
      "role": "support",
      "label": "AutoRound W4A16 INT4 · one user",
      "measurement_id": "qwen35-autoround-tp1-single-r16",
      "metric": "decode_tok_s",
      "sample_index": 0,
      "quality_label": "Experimental speed screen; 100/100 literal canaries, not a promoted deployment"
    },
    {
      "role": "support",
      "label": "AutoRound W4A16 INT4 · 64 users combined",
      "measurement_id": "qwen35-autoround-tp1-users-r16",
      "metric": "aggregate_tok_s",
      "sample_index": 0,
      "quality_label": "Experimental r16 two-point treatment; 100/100 literal canaries, but only 29/64 fixed-seed request sequences matched"
    },
    {
      "role": "support",
      "label": "Quark W8A8 INT8 · TP2 screen",
      "measurement_id": "qwen35-quark-tp2-screen",
      "metric": "decode_tok_s",
      "sample_index": 0,
      "quality_label": "Bounded speed screen; full quality suite was skipped"
    }
  ],
  "architecture": {
    "class": "Qwen3.6 35B A3B sparse MoE",
    "evidence": "results/qwen36-35b-quark-int8-b70/README.md"
  },
  "dimensions": {
    "weight_revision": [
      "qwen3.6-35b-a3b-base"
    ],
    "weight_quantization": [
      "Quark W8A8 INT8",
      "AutoRound W4A16 INT4"
    ],
    "runtime": [
      "vLLM XPU 2026-06 strict program",
      "vLLM XPU 0.20.2rc1.dev2+gc51df4300",
      "vLLM/XPU grouped-w4a16-piecewise-graph r14",
      "vLLM/XPU combined-runtime-guards r16"
    ],
    "tp": [
      1,
      2,
      4
    ],
    "mtp": [
      0,
      1,
      2,
      3,
      4
    ],
    "graph": [
      "on",
      "piecewise 1/32/64"
    ],
    "active_context_tokens": [
      0,
      2048,
      4096,
      8192,
      16384,
      24576,
      32768
    ],
    "configured_max_context_tokens": [
      4096,
      32768
    ],
    "simultaneous_users": [
      1,
      2,
      4,
      8,
      16,
      32,
      64
    ],
    "kv": [
      "auto"
    ]
  },
  "weight_revisions": [
    {
      "id": "qwen3.6-35b-a3b-base",
      "label": "Qwen3.6 35B A3B · shared base identity",
      "role": "conceptual base weight revision; retained evidence does not pin its upstream repository commit",
      "revision_status": "base commit unknown; only the quantized artifacts below carry artifact-specific provenance",
      "quantized_artifacts": [
        {
          "id": "qwen36-35b-quark-w8a8-cced565",
          "label": "Quark W8A8 INT8",
          "quantization": "Quark W8A8 INT8",
          "repository": "nameistoken/Qwen3.6-35B-A3B-Quark-W8A8-INT8",
          "revision": "cced56592e8c8935f8220836b4baa04dfd389118",
          "evidence": "data/localmaxxing-qwen36-35b-quark-int8-exacthf-20260612ak.json"
        },
        {
          "id": "qwen36-35b-autoround-w4a16",
          "label": "AutoRound W4A16 INT4",
          "quantization": "AutoRound W4A16 INT4",
          "repository": "abhinand/Qwen3.6-35B-A3B-int4-AutoRound",
          "revision_status": "Hugging Face revision was not captured in the retained experiment evidence",
          "evidence": "experiments/ornith-15-b70/data/2026-08-24-qwen36-35b-autoround-grouped-graph-aggregate-candidate.json"
        }
      ]
    }
  ],
  "model_variants": [],
  "transfer_scope": {
    "status": "one Qwen3.6 base weight revision; Quark and AutoRound are quantized deployment artifacts whose evidence stays separate",
    "transfers": [
      "the Qwen35MoE architecture boundary and general XPU graph investigation workflow",
      "the strict quality-gate method and archived failure-evidence discipline"
    ],
    "does_not_transfer": [
      "the legacy 99.428358 high into the newer deep-gate identity",
      "Quark speed, quality, topology, MTP failures, or lane closures into AutoRound",
      "AutoRound single-card speed or literal canaries into Quark or into any unmeasured TP, MTP, or context cell",
      "the invalid Quark MTP1 speed observation into a supported speculative result"
    ],
    "evidence": "results/qwen36-35b-quark-int8-b70/README.md"
  },
  "model_signals": {
    "b70_fit": {
      "band": "quant-specific: Quark TP4 strict; AutoRound TP1 screened",
      "scope": "Quark W8A8 INT8 and AutoRound W4A16 INT4 remain separate deployment artifacts of one base revision",
      "basis": "Quark completed a strict TP4 deep gate. AutoRound directly fit and ran on one B70 at a 4K configured ceiling, but remains an experimental determinism-open screen.",
      "reviewed_at": "2026-08-24"
    },
    "quality_evidence": {
      "band": "strict Quark TP4; bounded AutoRound TP1",
      "scope": "Quark TP4 passed the deep gate and Quark MTP1 failed immediately. AutoRound r16 passed 100/100 literal canaries, but fixed-seed B64 identity remains unresolved.",
      "evidence": [
        "data/qwen36-ablation-prefill-safe-int8-mixed-workspace-async-deep-gate-summary-20260615a13deep2.json",
        "data/qwen36-35b-autoround-b70-concurrency-20260824.json"
      ]
    },
    "popularity": {
      "state": "not-scored",
      "reason": "No dated popularity snapshot is stored for this exact revision."
    }
  },
  "run_measurements": [
    {
      "id": "qwen35-quark-tp4-strict-current",
      "state": "lab-measured",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-quark-w8a8-cced565",
      "variant": "Quark W8A8 INT8",
      "quantization": "Quark W8A8 INT8",
      "runtime": "vLLM XPU 2026-06 strict program",
      "runtime_identity": "prefill-safe INT8 mixed-workspace async deep gate 20260615a13deep2",
      "config": {
        "tp": 4,
        "mtp": 0,
        "graph": "on",
        "kv": "auto",
        "configured_max_context_tokens": 32768,
        "gate_generation": 2
      },
      "profile_id": "qwen35-quark-int8-tp4-current-strict-v1",
      "measurement_class": "strict deep gate",
      "promotion_status": "current strict reference",
      "quality_scope": "JSON 128/128, color 256/256, exact suite, repeat, long-context, and baseline match passed",
      "workload": "p512/o512, four post-warmup repeats, corrected output throughput after first streamed chunk",
      "metrics": {
        "decode_tok_s": [
          93.55054235558917
        ],
        "ttft_ms": [
          187.33663426246494
        ]
      },
      "evidence": "data/qwen36-ablation-prefill-safe-int8-mixed-workspace-async-deep-gate-summary-20260615a13deep2.json"
    },
    {
      "id": "qwen35-quark-tp4-legacy-approved",
      "state": "lab-measured",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-quark-w8a8-cced565",
      "variant": "Quark W8A8 INT8",
      "quantization": "Quark W8A8 INT8",
      "runtime": "vLLM XPU 0.20.2rc1.dev2+gc51df4300",
      "config": {
        "tp": 4,
        "mtp": 0,
        "graph": "on",
        "kv": "auto",
        "configured_max_context_tokens": 32768,
        "gate_generation": 1
      },
      "profile_id": "qwen35-quark-int8-tp4-legacy-approved-v1",
      "measurement_class": "older-gate LocalMaxxing-approved result",
      "promotion_status": "historical high",
      "quality_scope": "older exact/quality gates; retained as history, not substituted for the current strict reference",
      "workload": "p512/o512, four repeats after warmup, corrected steady-state output throughput after first streamed text chunk",
      "metrics": {
        "decode_tok_s": [
          99.42835812273452
        ]
      },
      "evidence": "data/localmaxxing-qwen36-35b-quark-int8-exacthf-20260612ak.json"
    },
    {
      "id": "qwen35-quark-tp2-screen",
      "state": "lab-screened",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-quark-w8a8-cced565",
      "variant": "Quark W8A8 INT8",
      "quantization": "Quark W8A8 INT8",
      "runtime": "vLLM XPU 2026-06 strict program",
      "config": {
        "tp": 2,
        "mtp": 0,
        "graph": "on",
        "kv": "auto",
        "configured_max_context_tokens": 32768
      },
      "profile_id": "qwen35-quark-int8-tp2-screen-v1",
      "measurement_class": "bounded smoke",
      "promotion_status": "screened only",
      "quality_scope": "JSON and color 16/16 passed; full quality suite was skipped",
      "workload": "p512/o512 TP2 smoke, one measured arm, corrected output throughput",
      "metrics": {
        "decode_tok_s": [
          85.86911405999231
        ]
      },
      "evidence": "data/qwen36-ablation-prefill-safe-int8-mixed-workspace-async-tp2-smoke-summary-20260615tp2safe1.json"
    },
    {
      "id": "qwen35-quark-tp4-mtp1-quarantined",
      "state": "quarantined",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-quark-w8a8-cced565",
      "variant": "Quark W8A8 INT8",
      "quantization": "Quark W8A8 INT8",
      "runtime": "vLLM XPU 2026-06 strict program",
      "config": {
        "tp": 4,
        "mtp": 1,
        "graph": "on",
        "kv": "auto",
        "configured_max_context_tokens": 32768
      },
      "profile_id": "qwen35-quark-int8-tp4-mtp1-parity-fix-v2",
      "measurement_class": "invalid speculative screen",
      "promotion_status": "quarantined",
      "quality_scope": "JSON and color failed on the first request; quality suite skipped",
      "workload": "p512/o512 MTP1 diagnostic, corrected output throughput; invalid for promotion",
      "metrics": {
        "decode_tok_s": [
          107.76565033909118
        ]
      },
      "evidence": "data/qwen36-ablation-tp4-mtp-k1-parity-fix-v2-summary-20260620061440.json"
    },
    {
      "id": "qwen35-autoround-tp1-single-r16",
      "state": "lab-screened",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-autoround-w4a16",
      "variant": "AutoRound W4A16 INT4",
      "quantization": "AutoRound W4A16 INT4",
      "runtime": "vLLM/XPU combined-runtime-guards r16",
      "config": {
        "tp": 1,
        "mtp": 0,
        "graph": "piecewise 1/32/64",
        "kv": "auto",
        "configured_max_context_tokens": 4096,
        "gpu_memory_utilization": 0.9,
        "natural_eos": false,
        "simultaneous_users": 1
      },
      "measurement_class": "experimental screen",
      "promotion_status": "experimental, not promoted",
      "quality_scope": "100/100 literal canaries across capture shapes; fixed-seed outputs identical 1/1 at one user",
      "workload": "i128/o1024, persistent raw-engine batching, single user, temperature 0, two repeats",
      "metrics": {
        "decode_tok_s": [
          90.90948338800597
        ]
      },
      "evidence": "data/qwen36-35b-autoround-b70-concurrency-20260824.json"
    },
    {
      "id": "qwen35-autoround-tp1-users-r16",
      "state": "lab-screened",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-autoround-w4a16",
      "variant": "AutoRound W4A16 INT4",
      "quantization": "AutoRound W4A16 INT4",
      "runtime": "vLLM/XPU combined-runtime-guards r16",
      "config": {
        "tp": 1,
        "mtp": 0,
        "graph": "piecewise 1/32/64",
        "kv": "auto",
        "configured_max_context_tokens": 4096,
        "gpu_memory_utilization": 0.9,
        "natural_eos": false,
        "simultaneous_users": 64
      },
      "measurement_class": "experimental discrete aggregate screen",
      "promotion_status": "experimental, not promoted; separate from the r14 curve",
      "quality_scope": "100/100 literal canaries across capture shapes; only 29/64 fixed-seed request sequences matched at 64 users",
      "workload": "i128/o1024 per request, persistent raw-engine batching, 64 simultaneous users, temperature 0, two repeats; combined counts every token after each request's first",
      "metrics": {
        "aggregate_tok_s": [
          1052.8704424119495
        ]
      },
      "evidence": "data/qwen36-35b-autoround-b70-concurrency-20260824.json"
    }
  ],
  "series_measurements": [
    {
      "id": "qwen35-autoround-tp1-users-r14",
      "state": "lab-screened",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-autoround-w4a16",
      "variant": "AutoRound W4A16 INT4",
      "quantization": "AutoRound W4A16 INT4",
      "runtime": "vLLM/XPU grouped-w4a16-piecewise-graph r14",
      "config": {
        "tp": 1,
        "mtp": 0,
        "graph": "piecewise 1/32/64",
        "kv": "auto",
        "configured_max_context_tokens": 4096,
        "gpu_memory_utilization": 0.9,
        "natural_eos": false
      },
      "measurement_class": "experimental screen",
      "promotion_status": "experimental, not promoted",
      "quality_scope": "4/4 literal canaries at four users; fixed-seed determinism at 64 users unresolved (27/64)",
      "workload": "i128/o1024 per request, persistent raw-engine batching, temperature 0, two repeats per level; combined counts every token after each request's first",
      "axis": "simultaneous_users",
      "points": [
        {
          "x": 1,
          "aggregate_tok_s": 87.6382088209863
        },
        {
          "x": 2,
          "aggregate_tok_s": 77.18628015630631
        },
        {
          "x": 4,
          "aggregate_tok_s": 149.48350787682995
        },
        {
          "x": 8,
          "aggregate_tok_s": 276.7655048798564
        },
        {
          "x": 16,
          "aggregate_tok_s": 488.6474098238277
        },
        {
          "x": 32,
          "aggregate_tok_s": 823.1508450629442
        },
        {
          "x": 64,
          "aggregate_tok_s": 1039.408012834992
        }
      ],
      "evidence": "data/qwen36-35b-autoround-b70-concurrency-20260824.json"
    }
  ],
  "estimates": [],
  "packets": [
    {
      "id": "qwen36-35b-quark-int8-b70-result",
      "label": "Qwen3.6 35B Quark INT8 · TP4 strict result collection",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-quark-w8a8-cced565",
      "quantization": "Quark W8A8 INT8",
      "runtime": "vLLM XPU 2026-06 strict program",
      "cards": 4,
      "status": "closed research result",
      "evidence_level": "B70 strict + historical records",
      "coverage": [
        "TP4 strict",
        "legacy high",
        "TP2 screen",
        "MTP1 quarantine",
        "quality"
      ],
      "grades": {
        "evidence": {
          "grade": "C",
          "scope": "Qwen3.6 35B Quark INT8 result collection",
          "basis": "strong TP4 deep-gate evidence and preserved history, but no clean-host package and only screened TP2 evidence; speculative paths remain invalid",
          "reviewed_at": "2026-08-24",
          "evidence": [
            "results/qwen36-35b-quark-int8-b70/README.md"
          ]
        }
      },
      "featured_metric": {
        "metric": "decode_tok_s",
        "measurement_id": "qwen35-quark-tp4-strict-current",
        "sample_index": 0,
        "value": 93.55054235558917,
        "unit": "tok/s",
        "workload": "p512/o512, four post-warmup repeats, corrected output throughput after first streamed chunk",
        "evidence": "data/qwen36-ablation-prefill-safe-int8-mixed-workspace-async-deep-gate-summary-20260615a13deep2.json"
      },
      "manifest": "results/qwen36-35b-quark-int8-b70/README.md"
    },
    {
      "id": "qwen36-35b-autoround-w4a16-research",
      "label": "Qwen3.6 35B AutoRound W4A16 · TP1 research evidence",
      "revision": "qwen3.6-35b-a3b-base",
      "artifact_id": "qwen36-35b-autoround-w4a16",
      "quantization": "AutoRound W4A16 INT4",
      "runtime": "vLLM/XPU combined-runtime-guards r16",
      "cards": 1,
      "status": "research",
      "evidence_level": "B70-measured experimental; determinism unresolved",
      "coverage": [
        "TP1",
        "MTP0",
        "4K configured ceiling",
        "one and 64-user screens",
        "literal canaries",
        "fixed-seed mismatch disclosed"
      ],
      "grades": {
        "evidence": {
          "grade": "C",
          "scope": "Qwen3.6 35B AutoRound W4A16 TP1 r14/r16 research program",
          "basis": "direct rates, repeat pairs, hashes, runtime identities, component probes, and literal canaries are retained, but the Hugging Face artifact revision was not captured, fixed-seed B64 identity remains unresolved, and there is no sealed deployment recipe",
          "reviewed_at": "2026-08-24",
          "evidence": [
            "docs/qwen36-35b-aggregate-throughput-evidence.md",
            "experiments/ornith-15-b70/data/2026-08-24-qwen36-35b-autoround-grouped-graph-aggregate-candidate.json"
          ]
        }
      },
      "featured_metric": {
        "metric": "decode_tok_s",
        "measurement_id": "qwen35-autoround-tp1-single-r16",
        "sample_index": 0,
        "value": 90.90948338800597,
        "unit": "tok/s",
        "workload": "i128/o1024, persistent raw-engine batching, single user, temperature 0, two repeats",
        "evidence": "data/qwen36-35b-autoround-b70-concurrency-20260824.json"
      },
      "manifest": "docs/qwen36-35b-aggregate-throughput-evidence.md"
    }
  ],
  "views": [
    {
      "id": "qwen35-tp4-history",
      "title": "TP4 strict and historical identities",
      "subtitle": "1 = older LocalMaxxing-approved 99.428358 high; 2 = current 93.550542 deep-gate reference. Different gate generations are discrete history, not a regression A/B.",
      "x_label": "gate generation",
      "discrete": true,
      "metrics": [
        "decode_tok_s",
        "ttft_ms"
      ],
      "series": [
        {
          "label": "older approved high",
          "measurement_ids": [
            "qwen35-quark-tp4-legacy-approved"
          ],
          "x_from": "config.gate_generation"
        },
        {
          "label": "current strict reference",
          "measurement_ids": [
            "qwen35-quark-tp4-strict-current"
          ],
          "x_from": "config.gate_generation"
        }
      ]
    },
    {
      "id": "qwen35-users-aggregate",
      "title": "One card serving many users",
      "subtitle": "Combined rate across all simultaneous users, one Arc Pro B70, AutoRound INT4 · experimental: the 2-16-user dip is the stack's graph-capture boundary, and 64-user determinism is still open",
      "x_label": "simultaneous users",
      "metrics": [
        "aggregate_tok_s"
      ],
      "series": [
        {
          "label": "measured combined rate",
          "measurement_ids": [
            "qwen35-autoround-tp1-users-r14"
          ]
        }
      ],
      "x_scale": "log"
    }
  ],
  "coverage_views": [
    {
      "id": "qwen35-current-mtp-by-tp",
      "label": "MTP × TP",
      "fixed": "Current June 2026 strict-program identity; the legacy TP4 high is shown separately.",
      "row_axis": {
        "key": "mtp",
        "label": "MTP",
        "prefix": "MTP"
      },
      "column_axis": {
        "key": "tp",
        "label": "TP",
        "prefix": "TP"
      },
      "fixed_selectors": {
        "revision": "qwen3.6-35b-a3b-base",
        "artifact_id": "qwen36-35b-quark-w8a8-cced565",
        "variant": "Quark W8A8 INT8",
        "runtime": "vLLM XPU 2026-06 strict program",
        "graph": "on",
        "kv": "auto",
        "configured_max_context_tokens": 32768
      },
      "rows": [
        0,
        1,
        2,
        3,
        4
      ],
      "columns": [
        1,
        2,
        4
      ],
      "cells": {
        "0:1": {
          "state": "missing",
          "label": "no stored lane"
        },
        "0:2": {
          "state": "lab-screened",
          "label": "85.869; full quality skipped",
          "evidence_id": "qwen35-quark-tp2-screen"
        },
        "0:4": {
          "state": "lab-measured",
          "label": "93.551 strict",
          "evidence_id": "qwen35-quark-tp4-strict-current",
          "packet_id": "qwen36-35b-quark-int8-b70-result"
        },
        "1:1": {
          "state": "missing",
          "label": "no stored lane"
        },
        "1:2": {
          "state": "missing",
          "label": "no stored lane"
        },
        "1:4": {
          "state": "quarantined",
          "label": "107.766 invalid",
          "evidence_id": "qwen35-quark-tp4-mtp1-quarantined"
        },
        "2:1": {
          "state": "missing",
          "label": "no Quark MTP2 arm"
        },
        "2:2": {
          "state": "missing",
          "label": "no Quark MTP2 arm"
        },
        "2:4": {
          "state": "missing",
          "label": "no Quark MTP2 arm"
        },
        "3:1": {
          "state": "missing",
          "label": "no Quark MTP3 arm"
        },
        "3:2": {
          "state": "missing",
          "label": "no Quark MTP3 arm"
        },
        "3:4": {
          "state": "missing",
          "label": "no Quark MTP3 arm"
        },
        "4:1": {
          "state": "missing",
          "label": "no Quark MTP4 arm"
        },
        "4:2": {
          "state": "missing",
          "label": "no Quark MTP4 arm"
        },
        "4:4": {
          "state": "missing",
          "label": "no Quark MTP4 arm"
        }
      }
    },
    {
      "id": "qwen35-quark-context-by-tp",
      "label": "Quark active context × TP",
      "fixed": "Canonical active prompt-depth grid. The retained Quark speed rows used p512/o512, and separate long-context quality checks are not relabeled as performance at any canonical depth.",
      "row_axis": {
        "key": "active_context_tokens",
        "label": "Active context",
        "prefix": "",
        "value_labels": {
          "0": "0",
          "2048": "2K",
          "4096": "4K",
          "8192": "8K",
          "16384": "16K",
          "24576": "24K",
          "32768": "32K"
        }
      },
      "column_axis": {
        "key": "tp",
        "label": "TP",
        "prefix": "TP"
      },
      "fixed_selectors": {
        "revision": "qwen3.6-35b-a3b-base",
        "artifact_id": "qwen36-35b-quark-w8a8-cced565",
        "variant": "Quark W8A8 INT8",
        "runtime_family": "vLLM XPU",
        "mtp": 0,
        "kv": "auto"
      },
      "rows": [
        0,
        2048,
        4096,
        8192,
        16384,
        24576,
        32768
      ],
      "columns": [
        1,
        2,
        4
      ],
      "cells": {
        "0:1": {"state": "missing", "label": "no exact zero-depth performance arm"},
        "0:2": {"state": "missing", "label": "no exact zero-depth performance arm"},
        "0:4": {"state": "missing", "label": "no exact zero-depth performance arm"},
        "2048:1": {"state": "missing", "label": "no exact 2K performance arm"},
        "2048:2": {"state": "missing", "label": "no exact 2K performance arm"},
        "2048:4": {"state": "missing", "label": "no exact 2K performance arm"},
        "4096:1": {"state": "missing", "label": "no exact 4K performance arm"},
        "4096:2": {"state": "missing", "label": "no exact 4K performance arm"},
        "4096:4": {"state": "missing", "label": "no exact 4K performance arm"},
        "8192:1": {"state": "missing", "label": "no exact 8K performance arm"},
        "8192:2": {"state": "missing", "label": "no exact 8K performance arm"},
        "8192:4": {"state": "missing", "label": "no exact 8K performance arm"},
        "16384:1": {"state": "missing", "label": "no exact 16K performance arm"},
        "16384:2": {"state": "missing", "label": "no exact 16K performance arm"},
        "16384:4": {"state": "missing", "label": "no exact 16K performance arm"},
        "24576:1": {"state": "missing", "label": "no exact 24K performance arm"},
        "24576:2": {"state": "missing", "label": "no exact 24K performance arm"},
        "24576:4": {"state": "missing", "label": "no exact 24K performance arm"},
        "32768:1": {"state": "missing", "label": "no exact 32K performance arm"},
        "32768:2": {"state": "missing", "label": "no exact 32K performance arm"},
        "32768:4": {"state": "missing", "label": "no exact 32K performance arm"}
      }
    },
    {
      "id": "qwen35-autoround-mtp-by-tp",
      "label": "AutoRound MTP × TP",
      "fixed": "AutoRound W4A16 r16, one-user p128/o1024 screen, 4K configured ceiling. Quark results and closures do not transfer into this slice.",
      "row_axis": {
        "key": "mtp",
        "label": "MTP",
        "prefix": "MTP"
      },
      "column_axis": {
        "key": "tp",
        "label": "TP",
        "prefix": "TP"
      },
      "fixed_selectors": {
        "revision": "qwen3.6-35b-a3b-base",
        "artifact_id": "qwen36-35b-autoround-w4a16",
        "variant": "AutoRound W4A16 INT4",
        "runtime": "vLLM/XPU combined-runtime-guards r16",
        "graph": "piecewise 1/32/64",
        "kv": "auto",
        "configured_max_context_tokens": 4096,
        "natural_eos": false,
        "simultaneous_users": 1
      },
      "rows": [
        0,
        1,
        2,
        3,
        4
      ],
      "columns": [
        1,
        2,
        4
      ],
      "cells": {
        "0:1": {
          "state": "lab-screened",
          "label": "90.909 experimental",
          "evidence_id": "qwen35-autoround-tp1-single-r16",
          "packet_id": "qwen36-35b-autoround-w4a16-research",
          "reason": "Direct speed and literal-canary evidence; not a deterministic deployment recipe."
        },
        "0:2": {
          "state": "missing",
          "label": "no AutoRound TP2 run"
        },
        "0:4": {
          "state": "missing",
          "label": "no AutoRound TP4 run"
        },
        "1:1": {
          "state": "missing",
          "label": "no AutoRound MTP1 arm"
        },
        "1:2": {
          "state": "missing",
          "label": "no AutoRound MTP1 arm"
        },
        "1:4": {
          "state": "missing",
          "label": "no AutoRound MTP1 arm"
        },
        "2:1": {
          "state": "missing",
          "label": "no AutoRound MTP2 arm"
        },
        "2:2": {
          "state": "missing",
          "label": "no AutoRound MTP2 arm"
        },
        "2:4": {
          "state": "missing",
          "label": "no AutoRound MTP2 arm"
        },
        "3:1": {
          "state": "missing",
          "label": "no AutoRound MTP3 arm"
        },
        "3:2": {
          "state": "missing",
          "label": "no AutoRound MTP3 arm"
        },
        "3:4": {
          "state": "missing",
          "label": "no AutoRound MTP3 arm"
        },
        "4:1": {
          "state": "missing",
          "label": "no AutoRound MTP4 arm"
        },
        "4:2": {
          "state": "missing",
          "label": "no AutoRound MTP4 arm"
        },
        "4:4": {
          "state": "missing",
          "label": "no AutoRound MTP4 arm"
        }
      }
    },
    {
      "id": "qwen35-autoround-context-by-tp",
      "label": "AutoRound active context × TP",
      "fixed": "Canonical active prompt-depth grid. The retained AutoRound run used p128/o1024 with max_model_len=4K, so it is not relabeled as a 0, 2K, or 4K active-context point.",
      "row_axis": {
        "key": "active_context_tokens",
        "label": "Active context",
        "prefix": "",
        "value_labels": {
          "0": "0",
          "2048": "2K",
          "4096": "4K",
          "8192": "8K",
          "16384": "16K",
          "24576": "24K",
          "32768": "32K"
        }
      },
      "column_axis": {
        "key": "tp",
        "label": "TP",
        "prefix": "TP"
      },
      "fixed_selectors": {
        "revision": "qwen3.6-35b-a3b-base",
        "artifact_id": "qwen36-35b-autoround-w4a16",
        "variant": "AutoRound W4A16 INT4",
        "runtime_family": "vLLM/XPU",
        "mtp": 0,
        "kv": "auto"
      },
      "rows": [
        0,
        2048,
        4096,
        8192,
        16384,
        24576,
        32768
      ],
      "columns": [
        1,
        2,
        4
      ],
      "cells": {
        "0:1": {"state": "missing", "label": "no exact zero-depth arm"},
        "0:2": {"state": "missing", "label": "no exact zero-depth arm"},
        "0:4": {"state": "missing", "label": "no exact zero-depth arm"},
        "2048:1": {"state": "missing", "label": "no exact 2K active-depth arm"},
        "2048:2": {"state": "missing", "label": "no exact 2K active-depth arm"},
        "2048:4": {"state": "missing", "label": "no exact 2K active-depth arm"},
        "4096:1": {"state": "missing", "label": "no exact 4K active-depth arm"},
        "4096:2": {"state": "missing", "label": "no exact 4K active-depth arm"},
        "4096:4": {"state": "missing", "label": "no exact 4K active-depth arm"},
        "8192:1": {"state": "missing", "label": "no exact 8K active-depth arm"},
        "8192:2": {"state": "missing", "label": "no exact 8K active-depth arm"},
        "8192:4": {"state": "missing", "label": "no exact 8K active-depth arm"},
        "16384:1": {"state": "missing", "label": "no exact 16K active-depth arm"},
        "16384:2": {"state": "missing", "label": "no exact 16K active-depth arm"},
        "16384:4": {"state": "missing", "label": "no exact 16K active-depth arm"},
        "24576:1": {"state": "missing", "label": "no exact 24K active-depth arm"},
        "24576:2": {"state": "missing", "label": "no exact 24K active-depth arm"},
        "24576:4": {"state": "missing", "label": "no exact 24K active-depth arm"},
        "32768:1": {"state": "missing", "label": "no exact 32K active-depth arm"},
        "32768:2": {"state": "missing", "label": "no exact 32K active-depth arm"},
        "32768:4": {"state": "missing", "label": "no exact 32K active-depth arm"}
      }
    }
  ],
  "family_closures": [
    {
      "selectors": {
        "revision": "qwen3.6-35b-a3b-base",
        "artifact_id": "qwen36-35b-quark-w8a8-cced565",
        "variant": "Quark W8A8 INT8",
        "frontier": "ad-hoc-optimization"
      },
      "state": "closed",
      "reason": "The TP4 target-only lane is exhausted for now; reopen only for a controlled upstream bakeoff or a new graph-compatible speculative-state mechanism.",
      "evidence": "results/qwen36-35b-quark-int8-b70/README.md"
    }
  ]
}