language: [en]
license: apache-2.0
pipeline_tag: text-generation
base_model: Qwen/Qwen3.8-27B
base_model_relation: quantized
quantized_by: mlasli
library_name: mlx
tags:

• qwen3.8
• qwen3
• abliterated
• heretic
• uncensored
• decensored
• text-generation
• roleplay
• mlx
• quantization
• apple-silicon

Qwen3.8-27B Heretic-Abliterated (8-bit MLX)

MLX quantization of
mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16
— Qwen/Qwen3.8-27B with its refusal direction removed using
Heretic (single-direction abliteration with an
Optuna-based parameter search). The language backbone is abliterated; all other
capabilities are preserved.

What this is for: the same abliterated 27B weights, quantized to 28.6 GB in MLX format
(group-wise 8-bit, group size 64) for fast Apple-Silicon inference via mlx-lm.

Text-only. The MLX conversion drops the vision tower (only language_model.* weights are
shipped), so this repo handles text only. For image input use the BF16 safetensors repo
or the GGUF repos with a separate mmproj.

Note on the parameter count badge: this is a 27B-parameter model. Hugging
Face's automatic scanner may show a lower count (e.g. "6B" / "8B") for MLX
quantizations because it counts the packed 32-bit integer elements (which hold
multiple low-bit weights) as individual parameters rather than unpacking them.
The true total in model.safetensors.index.json is ~27B.

Note on the parameter count badge: this is a 27B-parameter model. Hugging
Face's automatic scanner may show a lower count (e.g. "6B" / "8B") for MLX
quantizations because it counts the packed 32-bit integer elements (which hold
multiple low-bit weights) as individual parameters rather than unpacking them.
The true total in model.safetensors.index.json is ~27B.

What is Heretic?

Heretic is an abliteration method that removes a
model's safety-aligned refusal direction in one shot (unlike earlier multi-direction
approaches), trading a minimal amount of capability for a large drop in refusals. Its
Optuna search picks the ablation parameters on the Pareto front of
(compliance, first-token KL divergence).

Evaluation

Source: Independent eval (merged model)

• Compliance: 94.0% (harmful-behaviors, Zou et al. refusal detector, 50 prompts)
• Zou 29-substring refusal rate: 6.0%
• First-token KL divergence vs base: 0.0467

A stricter combined keyword detector reported 18.0% refusal, but manual review of the
flagged completions confirmed these are largely false positives (the model answers
directly and uses words like "illegal"/"harmful"/"violent" inside compliant responses).
The Zou number above is the more reliable refusal estimate.

Usage

pip install -U mlx-lm

# one-shot generation
mlx_lm.generate \
  --model mlasli/Qwen3.8-27B-Heretic-Abliterated-MLX-8bit \
  --prompt "Hi!" --max-tokens 256

# OpenAI-compatible server
mlx_lm.server \
  --model mlasli/Qwen3.8-27B-Heretic-Abliterated-MLX-8bit --port 8080

Quantizations

MLX quants (group-wise 6/8-bit, group size 64 — not GGUF Q6_K/Q8_0):

GGUF quants (llama.cpp / Ollama) are in separate repos:

Abliteration removes safety alignment. Use responsibly and in accordance with your local
laws and the upstream Apache-2.0 license.

Changelog

v1.0.0 — initial MLX release (2026-08-16)

• Initial MLX quantization of mlasli/Qwen3.8-27B-Heretic-Uncensored-BF16.
• Group-wise 8-bit (group size 64) affine quantization via mlx_lm.convert.