Modelpublic

philbert440/Qwen3.8-27B-Uncensored-Aggressive-MLX-4bit

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes409downloads

Model Card

Qwen3.8-27B-Uncensored-Aggressive-MLX-4bit

4-bit MLX quantization of philbert440/Qwen3.8-27B-Uncensored-Aggressive for Apple Silicon (MLX). Group size 64, affine 4-bit.

Aggressive single-direction refusal ablation of Qwen3.8-27B (Heretic/Arditi lineage) — broadly de-refused while retaining reasoning and factual honesty. See the base model card for recipe and evaluation details.

Run

bash

pip install mlx-lm
mlx_lm.generate --model philbert440/Qwen3.8-27B-Uncensored-Aggressive-MLX-4bit --prompt "..."

Requires mlx-lm >= 0.31.3 (adds the qwen3_5 architecture).

Note

Uncensored model — you are responsible for how you use it. Thinking-mode use benefits from an ample token budget and sampling (temp ~0.7); greedy/temperature-0 can loop on long open-ended prompts.