Qwen3.8-27B Uncensored

🔬 Research build · refusal-removed via abliteration

A fully unlocked 27B hybrid-attention model, run on your own machine

Gated DeltaNet linear attention + full attention, native vision, 262K context, and zero refusals — on Apple Silicon via MLX. Pick a quant, copy a command, run it.

27B dense 262K context Hybrid Gated DeltaNet MTP head Vision + Tools 0 refusals Apache 2.0

Run the MLX build on Apple Silicon

pip install -U mlx-vlm # needs mlx-vlm ≥ 0.6.13, mlx ≥ 0.32

hf download orcarouter/Qwen3.8-27B-Uncensored-MLX --include "4-bit/*" --local-dir ./qwen-unc

python -m mlx_vlm server --model ./qwen-unc/4-bit --port 8080

The repo is gated — sign in on Hugging Face, accept the terms, then hf auth login. The server exposes an OpenAI-compatible endpoint at http://localhost:8080/v1. Vision: add --image path/to/image.png to mlx_vlm generate.

💡 RAM guidance: 4-bit ≈ 15 GB (24 GB Macs) · 6-bit ≈ 22 GB (24–32 GB) · 8-bit ≈ 27.5 GB (32 GB, near-lossless). The vision tower stays BF16 in every build, so leave a few GB of headroom for it + context.

⚠️ This model has had its safety alignment substantially removed. Released strictly for legitimate research (interpretability, AI-safety and refusal-mechanism study, red-teaming, robustness evaluation, controlled experiments). You assume full responsibility and liability for use. Add your own moderation before any deployment. Apache 2.0 license inherited from the base model.

MLX quantizations — pick your precision

Only the language-model linear layers are quantized (affine, group size 64). The vision tower, norms and conv layers stay BF16. All builds verified by generation on GPU; refusal probes return 0 refusals on 4/6/8-bit.

# select a quant above to see the exact download + run commands

Architecture — 64 layers, hybrid attention

Gated DeltaNet linear attention (48 layers) Full attention every 4th layer (16 layers)

hidden 5120 · 24 attn heads · 4 KV heads · 17408 intermediate · conv kernel 4 · MTP (1 layer) · 248,320 vocab

⚠️ Uncensored model. Safety alignment has been removed via abliteration. It will comply with harmful, unethical, offensive or illegal requests that Qwen3.8-27B would refuse, and can produce confident falsehoods, bias and offensive text. Intended strictly for AI-safety research, interpretability study, red-teaming and controlled experiments. You assume full responsibility for use. Not for deployment to end users without your own moderation layer.