Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-mlx-8bit

MLX-VLM conversion of huihui-ai/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated.

Overview

• Format: MLX-VLM
• Precision: 8bit
• Size: about 35G
• Quantization result: 8.596 bits/weight
• Source model is preserved as a vision-language model for mlx-vlm
• Local validation passed for text generation and abliterated behavior regression

Conversion Notes

This conversion keeps the model in the mlx-vlm layout and includes the compatibility fixes required for reliable use with MLX-VLM and LM Studio:

• restored a Qwen VL-compatible chat_template.jinja
• aligned bos/eos/pad token ids in config.json
• preserved image and video prompt token handling

Validation

Local checks performed on Apple Silicon:

• text generation smoke test: passed
• abliterated regression set: 6/6 non-refused
• refusal_rate = 0.0
• eval run id: 20260317_200037
• median cleaned response length: 546 chars
• eval settings: max_tokens=320, temperature=0.0, prefill_step_size=128

This is a behavior regression check, not a mathematical proof of equivalence.

Files

Important files in this repo:

• config.json
• chat_template.jinja
• processor_config.json
• tokenizer.json
• model-00001-of-00008.safetensors ... model-00008-of-00008.safetensors
• model.safetensors.index.json

Usage

mlx-vlm text generation

mlx_vlm.generate \
  --model /path/to/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-mlx-8bit \
  --prompt "你好" \
  --max-tokens 256 \
  --prefill-step-size 128

mlx-vlm image prompt

mlx_vlm.generate \
  --model /path/to/Huihui-Qwen3.5-35B-A3B-Claude-4.6-Opus-abliterated-mlx-8bit \
  --image /path/to/example.png \
  --prompt "请简短描述这张图片。" \
  --max-tokens 128 \
  --prefill-step-size 128

LM Studio

This repo is intended to work as an MLX model in LM Studio after download or sync. The included chat template already contains the required Qwen vision tokens.