library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
base_model:
- Qwen/Qwen3.5-2B
tags: - abliterated
- uncensored
- mlx
Huihui-Qwen3.5-2B-abliterated — MXFP8 MLX
MLX-converted version of huihui-ai/Huihui-Qwen3.5-2B-abliterated for Apple Silicon, quantized with MXFP8.
Model Details
| Property | Value |
|---|---|
| Base model | Qwen3.5-2B |
| Type | Vision-Language Model (VLM) |
| Format | MLX MXFP8 (~9.4 bits/weight) |
| Size | ~2.6 GB |
| Abliterated | Yes — censorship layers removed |
MXFP8 provides near-native quality with ~2x compression vs fp16 — better than traditional 8-bit quantization (lower perplexity, higher benchmark scores).
Variants
| Variant | Size | Quality | Link |
|---|---|---|---|
| fp16 | ~4.5 GB | Highest | fp16 |
| MXFP8 | ~2.6 GB | Near-native | This repo |
| MXFP4 | ~1.7 GB | Good | mxfp4 |
Usage
pip install mlx-vlm
# Text generation
python -m mlx_vlm.generate \
--model AITRADER/Huihui-Qwen3.5-2B-abliterated-mxfp8-MLX \
--prompt "Describe this image in detail" \
--image <path-or-url>
# Chat UI
python -m mlx_vlm.chat_ui \
--model AITRADER/Huihui-Qwen3.5-2B-abliterated-mxfp8-MLX