orcarouter / Qwen3.8-27B-Uncensored-MLX
27B vision model,
Apple Silicon first.
Demo for the MLX build of Qwen3.8-27B. Cloud chat uses official Qwen/Qwen3.8-27B via Hugging Face Inference Providers — the same architecture, aligned weights. Exact uncensored MLX quants stay on your Mac.
Research-use warning: the MLX artifact has had safety alignment substantially removed.
Do not treat outputs as advice, and do not deploy uncensored weights without moderation.
Exact MLX weights on Apple Silicon
Repo root is the recommended 4-bit build (~15 GB; plan for 24 GB unified memory). Needs mlx-vlm >= 0.6.13.
python -m pip install -U "mlx-vlm>=0.6.13" "mlx>=0.32" hf download orcarouter/Qwen3.8-27B-Uncensored-MLX \ --include "4-bit/*" \ --local-dir ./Qwen3.8-27B-Uncensored-MLX python -m mlx_vlm generate \ --model ./Qwen3.8-27B-Uncensored-MLX/4-bit \ --prompt "Explain quantum entanglement in one sentence." \ --max-tokens 256
Vision: add --image path/to/image.png. Local OpenAI-compatible server:
python -m mlx_vlm server \ --model ./Qwen3.8-27B-Uncensored-MLX/4-bit \ --port 8080
Optional MTP speculative decoding (any precision + the mtp/ drafter):
hf download orcarouter/Qwen3.8-27B-Uncensored-MLX --include "6-bit/*" "mtp/*" \ --local-dir ./Qwen3.8-27B-Uncensored-MLX python -m mlx_vlm generate \ --model ./Qwen3.8-27B-Uncensored-MLX/6-bit \ --draft-model ./Qwen3.8-27B-Uncensored-MLX/mtp \ --draft-kind mtp --draft-block-size 4 \ --prompt "Explain quantum entanglement in one sentence." \ --max-tokens 256
| Build | Size | Mac memory |
|---|---|---|
| 4-bit | ~15 GB | 24 GB+ recommended |
| 6-bit | ~22 GB | 24–32 GB+ |
| 8-bit | ~27.5 GB | 32 GB+ |
| 2-bit | ~8.7 GB | archival only |
What this Space serves
- Requested artifact: orcarouter/Qwen3.8-27B-Uncensored-MLX — MLX affine quants of abliterated Qwen3.8-27B.
- Cloud playground: official Qwen/Qwen3.8-27B on Inference Providers. A free Hugging Face account cannot host Gradio/ZeroGPU yet, so the 54 GB uncensored FP8 companion is not loaded here.
- Architecture: 27B dense hybrid attention (Gated DeltaNet + full attention), native vision tower, 262K context, thinking control, tool-calling, MTP head.
Apache-2.0, inherited from Qwen. Intended for controlled research — interpretability, refusal study, and robustness evaluation.