Runtime target: stock mlx-vlm and LM Studio MLX engine.
Custom runtime required: no.
Vision preserved: yes.
Video processor preserved: yes.
MTP keys: 0 keys; source has 19. Current mlx-vlm qwen3_5_moe sanitizer drops mtp.* keys, so MTP runtime support is not claimed..
MTP runtime support: not claimed unless stated in the validation table.

Recipe

json

{
  "profile": "VLM-8bit",
  "convert_status": "converted",
  "convert_reason": null,
  "converter": "mlx_vlm.convert",
  "source_model": "WaveCut/Qwen3.6-35B-A3B-REAM-160-ru-agent",
  "source_revision": "c3b73b8aaf70e2340d9b16424741213b5743936d",
  "output_format": "MLX safetensors with mlx-vlm processor files",
  "quantization_summary": {
    "present": true,
    "global": {
      "group_size": 64,
      "bits": 8,
      "mode": "affine"
    },
    "explicit_rule_count": 80,
    "rules_by_bits_group_mode": [
      {
        "bits": "8",
        "group_size": "64",
        "mode": "affine",
        "count": 80
      }
    ],
    "rules_by_path_class": {
      "router_or_gate": 80
    }
  }
}

Validation

check	result
`mlx_vlm.load`	`yes`
text smoke	`yes`
image smoke	`yes`
video smoke	`skipped unless supported by current runtime`
LM Studio smoke	`not tested yet`

Quality

json

{
  "metric": "proxy_smoke_score_not_kl",
  "kl_status": "not_computed",
  "domain_scores": {
    "ru": 0.7751475733057623,
    "agent": 0.9,
    "code": 0.65,
    "math": 0.8,
    "vision": 0.8999999999999999
  },
  "weighted_score": 0.8325442719917286,
  "text_kl": {
    "status": "computed",
    "summary": {
      "mean": 0.0008244836847097758,
      "p95": 0.0025505936238914718,
      "max": 0.003115440718829632,
      "weighted_mean": 0.0008720117882856154,
      "same_top": 1.0,
      "domains": {
        "agent": {
          "mean": 7.906566406745696e-05,
          "p95": 0.0001360272567580978,
          "max": 0.00014235632261261344,
          "count": 2,
          "same_top": 1.0
        },
        "code": {
          "mean": 0.000385318482585717,
          "p95": 0.0007160460583691019,
          "max": 0.0007527935667894781,
          "count": 2,
          "same_top": 1.0
        },
        "math": {
          "mean": 0.0008301898196805269,
          "p95": 0.0014344516705023123,
          "max": 0.0015015918761491776,
          "count": 2,
          "same_top": 1.0
        },
        "ru": {
          "mean": 0.0020033607725054026,
          "p95": 0.0030042327241972086,
          "max": 0.003115440718829632,
          "count": 2,
          "same_top": 1.0
        }
      }
    }
  },
  "mmlu_pro_validation20": {
    "status": "not_computed",
    "dataset": null,
    "split": null,
    "sample_count": null,
    "summary": null,
    "method": null,
    "official_leaderboard": false
  },
  "mmlu_pro_test1000": {
    "status": "not_computed",
    "dataset": null,
    "split": null,
    "sample_count": null,
    "summary": null,
    "method": null,
    "official_leaderboard": false,
    "complete": null
  },
  "comparison_note": "Ranking uses text KL as the primary quantization signal, adds 1000-question MMLU-Pro test as a quality tie-breaker when available, and falls back to proxy smoke quality when KL is unavailable.",
  "comparison_row": {
    "profile": "VLM-8bit",
    "converted": true,
    "convert_status": "converted",
    "convert_reason": null,
    "structural_ok": true,
    "smoke_ok": true,
    "quality_score": 0.8325442719917286,
    "kl_status": "computed",
    "kl_weighted_mean": 0.0008720117882856154,
    "kl_mean": 0.0008244836847097758,
    "same_top": 1.0,
    "mmlu_pro_validation20_accuracy": null,
    "mmlu_pro_validation20_mean_margin": null,
    "mmlu_pro_test1000_accuracy": null,
    "mmlu_pro_test1000_mean_margin": null,
    "size_bytes": 24925132285,
    "score": 3.266995
  },
  "lm_studio": {
    "status": "not_computed",
    "ok": null,
    "text_chat_ok": null,
    "image_chat_ok": null,
    "mtp_status": null,
    "note": null
  }
}

Reproducibility

Source model: WaveCut/Qwen3.6-35B-A3B-REAM-160-ru-agent
Source revision: c3b73b8aaf70e2340d9b16424741213b5743936d
Tool versions: see manifests/tool_versions.json.
Full profile manifest: see manifests/profile-VLM-8bit.json.