Qwen3.8-27B-MLX-3bit-Group32

Qwen3.8-27B MLX-3bit-Group32

This is a vanilla quantization of Qwen/Qwen3.8-27B. It is not a fine-tune, merge, ablation, alignment change, or chat-template modification. The source weights are pinned to commit 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

The official checkpoint uses Qwen3_5ForConditionalGeneration / qwen3_5 as its internal architecture identifier. That string does not mean these weights came from a Qwen3.5 model.

Conversion

{
  "algorithm": "MLX affine 3-bit quantization with group size 32",
  "bit_width": 3,
  "group_size": 32,
  "calibration_source": "none"
}

• Source tensor inventory: 1199 tensors, including 333 vision tensors and 15 source MTP tensors.
• Conversion tool/runtime requirement: mlx-vlm / 0.6.1.
• Artifact size: 15.266 GB (decimal).
• Expected hardware: Apple Silicon with at least 32 GB unified memory.

Calibration source: none.

Component status

• Text: passed release tests.
• Vision/video: passed deterministic local image tests.
• Tool calling: passed all native XML tool tests.
• MTP: loaded and passed a temperature-zero equivalence and throughput A/B.
• Chat template, tokenizer, processor, generation config, and special-token IDs: checked against the locked source by the structural gate.
• Quality comparison: passed against the locked BF16 source using the exact same functional cases. Semantic similarity uses sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 at e8f8c211226b894fcb81acc59f3b34ba3efd5f42 as a measured proxy, not as ground-truth accuracy.
• Longest recorded validation prompt: 73 prompt tokens. This is a measured test boundary, not a claim that the architectural maximum was exercised.

Validation results

{
  "release_gate": "PASS",
  "text": [
    true,
    true,
    true,
    true,
    true,
    true,
    true,
    true,
    true,
    true
  ],
  "tools": [
    true,
    true,
    true,
    true,
    true
  ],
  "vision": [
    true,
    true,
    true
  ],
  "mtp": {
    "passed": true,
    "drafter_kind": "mtp",
    "output_equivalent_temperature_zero": true,
    "accepted_drafts": 34,
    "drafted_tokens": 34,
    "acceptance_rate": 1.0,
    "baseline_tps": 17.784681920889266,
    "mtp_tps": 15.948037679600025,
    "speedup": 0.8967288676030813,
    "measured_improvement": false,
    "baseline_wall_seconds": 3.0863686250522733,
    "mtp_wall_seconds": 3.355042084120214,
    "advertise_acceleration": false
  },
  "bf16_source_comparison": {
    "passed": true,
    "mean_semantic_similarity": 0.9414990961551666,
    "exact_matches": 3,
    "measurements": {
      "average_generation_tps": 18.347696131763715,
      "peak_memory_gb": 15.917112695,
      "artifact_bytes": 15265823142,
      "maximum_prompt_tokens_tested": 73,
      "loop_rate": 0.0
    },
    "evaluator": {
      "repo_id": "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2",
      "revision": "e8f8c211226b894fcb81acc59f3b34ba3efd5f42",
      "pooling": "attention-mask mean pooling followed by L2 normalization",
      "maximum_tokens": 256
    }
  },
  "bf16_fixed_logit_comparison": {
    "positions": 106,
    "inputs_sha256": "f1a0af6b9580739ebc9efa9375ae31aa6940dafe7ff6446097ed6cf8d9ab37de",
    "mean_kl_divergence": 0.09177716048258655,
    "reference_perplexity": 9.84727010437735,
    "candidate_perplexity": 10.580787393174923,
    "perplexity_delta": 0.7335172887975734,
    "top1_token_agreement": 0.8301886792452831,
    "selection": "selected",
    "warnings": [
      "KL is measured on fixed original text, not a public benchmark.",
      "BF16 log-probabilities are stored in float16 after float32 log-softmax; reported KL therefore has finite-storage approximation error."
    ]
  }
}

No acceleration is advertised unless the MTP report contains a measured throughput improvement. Exact measurements are artifact-, prompt-, context-, and hardware-specific.

Inference

python -m pip install 'mlx==0.31.2' 'mlx-lm==0.31.3' 'mlx-vlm==0.6.1' 'huggingface-hub[cli]'
hf download Chungulus/Qwen3.8-27B-MLX-3bit-Group32 --local-dir ./qwen38-quant
python -m mlx_vlm.generate --model ./qwen38-quant --draft-model ./qwen38-quant/mtp-drafter --draft-kind mtp --draft-block-size 3 --prompt 'Describe this image.' --image ./image.png --max-tokens 256 --no-verbose

Use the exact source chat-template controls for thinking (enable_thinking, reasoning_effort, and preserve_thinking) and the native Qwen tool format.

Limitations

Quantization can reduce quality, especially at very low bit widths. Runtime support for the hybrid Gated DeltaNet/full-attention graph, vision tower, projector, processor, and MTP component is format-specific. A loader that reads only a language tensor is not sufficient. Tested context length and resource measurements are recorded in validation_result.json; untested context lengths must not be inferred from the architectural maximum.

License and attribution

The parent model and this unmodified quantization are distributed under the source model's Apache-2.0 license. See the official Qwen3.8-27B repository for the upstream model card and attribution.