vanch007/Huihui-Qwen3.5-9B-abliterated-mlx-4bit

Huihui-Qwen3.5-9B-abliterated-mlx-4bit

MLX-VLM 4bit export of huihui-ai/Huihui-Qwen3.5-9B-abliterated for Apple Silicon workflows, including LM Studio and local mlx_vlm usage.

Overview

• —Variant: 4bit
• —Repository payload at upload time: 6.0G
• —Repository file count: 13
• —Effective quantization observed during conversion: 5.059 bits per weight
• —Format: mlx-vlm model package

Compatibility

• —Uses the corrected Qwen VL chat template with image token placeholders.
• —Uses <|im_end|>-compatible stop token settings for cleaner chat termination in MLX/LM Studio.

Validation

• —Local text generation smoke test: passed
• —Local image-input smoke test: passed
• —Local black-box abliterated check: 6/6 non-refused
• —Refusal rate: 0.0
• —Actionable non-refused cases: 6
• —Median cleaned response length: 1263 chars

Behavior Notes

• —This variant preserved the abliterated behavior on the local 6-case regression set used during conversion validation.
• —These checks are behavioral acceptance tests, not a formal guarantee of identical outputs to the source checkpoint.

Usage

mlx_vlm.generate \
  --model /path/to/Huihui-Qwen3.5-9B-abliterated-mlx-4bit \
  --prompt "你好" \
  --max-tokens 256

For behavior-focused checks, it is safer to disable thinking output so refusal scoring is based on the final answer instead of the thinking trace.

mlx_vlm.generate \
  --model /path/to/Huihui-Qwen3.5-9B-abliterated-mlx-4bit \
  --prompt "你好" \
  --max-tokens 256 \
  --processor-kwargs '{"enable_thinking": false}'

LM Studio

• —If LM Studio has an older cached copy, refresh or re-download the repository so the latest chat template and config are picked up.
• —These repositories are meant for mlx-vlm / Apple MLX runtimes rather than Transformers CPU inference.