Qwen3.8-27B JANG_4S

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": "official JANG_4S adaptive mixed-precision MSE quantization",
  "bit_width": "approximately_4.1_average",
  "group_size": "converter_auto",
  "calibration_source": "none; official JANG MSE weight-only conversion"
}

• Source tensor inventory: 1199 tensors, including 333 vision tensors and 15 source MTP tensors.
• Conversion tool/runtime requirement: JANG/MLX / 2.5.46.
• Artifact size: 17.081 GB (decimal).
• Expected hardware: Apple Silicon.

Calibration source: none; official JANG MSE weight-only conversion.

Component status

• Text: passed release tests.
• Vision/video: passed deterministic local image tests.
• Tool calling: passed all native XML tool tests.
• MTP: source MTP tensors were retained by the structural gate; this repository does not claim speculative acceleration.
• 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,
    "preserved": true,
    "tensor_count": 31,
    "metadata_mode": "preserved_enabled",
    "runtime_supported": false,
    "measurement_required": false,
    "acceptance_rate": null,
    "baseline_tps": null,
    "mtp_tps": null,
    "speedup": null,
    "measured_improvement": false,
    "advertise_acceleration": false,
    "limitation": "JANG 2.5.46 preserves the Qwen3.8 MTP tensors, but its public mlx-vlm loader filters them because the model class has no native MTP module; no acceptance-rate or speed A/B can be measured."
  },
  "bf16_source_comparison": {
    "passed": true,
    "mean_semantic_similarity": 0.9505505561828613,
    "exact_matches": 4,
    "measurements": {
      "average_generation_tps": 17.21587225265724,
      "peak_memory_gb": 19.511131463,
      "artifact_bytes": 17080508919,
      "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
    }
  }
}

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 vllm huggingface-hub
vllm serve Chungulus/Qwen3.8-27B-JANG_4S

Use JANG/MLX at or above 2.5.46; the exact validated container digest is recorded in quantization_manifest.json.

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.