Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64

MTPLX / Apple Silicon optimized Q6G64 conversion of:

petruhonk/Qwen3.8-9B-Distill-uncensored-heretic

Repository:

xy10086/Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64

Highlights

• Qwen3.8 9B multimodal architecture
• Q6 affine quantization, group size 64
• Native BF16 MTP head preserved
• 15/15 MTP tensors
• BF16 vision tower preserved
• 333 vision tensors
• Image input verified
• MTP speculative decoding verified
• Recommended MTP depth: 2
• Tested on Apple M5 Pro, 48 GB unified memory

Provenance

Source model:

petruhonk/Qwen3.8-9B-Distill-uncensored-heretic

Upstream model:

empero-ai/Qwen3.8-9B-Distill

The source repository did not include the vision preprocessing sidecars required by MTPLX. Matching preprocessor_config.json and video_preprocessor_config.json were restored from the upstream empero-ai/Qwen3.8-9B-Distill repository.

The source vision weights were preserved by MTPLX Forge and stored in model-vision.safetensors.

Quantization

Main language-model trunk:

• 6-bit
• Group size 64
• Affine quantization

MTP:

• BF16
• 15 tensors
• Not requantized

Vision:

• BF16
• 333 tensors
• About 870 MiB on disk

Forge recipe: forge-recipe.json

Tested software

• MTPLX 2.11.2
• MLX 0.32.2
• mlx-lm 0.31.3

Recommended configuration

• Profile: sustained
• MTP depth: 2
• Context window: 32768
• Max response tokens: 8192

Benchmarks

MTPLX Forge verification

Real Chinese chat

Vision test

• 869 prompt tokens
• 336 completion tokens
• 65.30 decode tok/s
• Successful screenshot understanding and OCR
• Approx. 10.53 GB peak runtime memory

See benchmarks.json for the machine-readable summary.

Usage

Pull:

mtplx pull xy10086/Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64

Serve:

mtplx serve \
  --model xy10086/Qwen3.8-9B-Distill-Uncensored-Heretic-MTPLX-Q6G64 \
  --host 127.0.0.1 \
  --port 8000 \
  --profile sustained \
  --generation-mode mtp \
  --depth 2 \
  --context-window 32768 \
  --max-tokens 8192 \
  --reasoning off \
  --reasoning-effort low \
  --no-auth \
  --yes

OpenAI-compatible endpoint:

http://127.0.0.1:8000/v1

Behavior

This model is derived from a low-refusal / censorship-ablated checkpoint. Lower refusal behavior does not imply higher factual accuracy.

Medical, legal, financial and other high-stakes outputs should be independently verified.

Credits

• petruhonk/Qwen3.8-9B-Distill-uncensored-heretic
• empero-ai/Qwen3.8-9B-Distill
• MTPLX

License

Apache-2.0. Review the upstream repositories for their attribution and license requirements as well.