base_model: huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated
base_model_relation: quantized
license: apache-2.0
language:

• en
  library_name: mlx
  pipeline_tag: image-text-to-text
  tags:
• abliterated
• uncensored
• qwen3-vl
• qwen3
• vision-language
• multimodal
• mlx
• mlx-4bit
• apple-silicon
• on-device
• image-text-to-text
• mlx-lm
• local-llm
• macbook
• 32b
• 4-bit

Huihui-Qwen3-VL-32B-Instruct-abliterated-4bit-mlx

Uncensored vision-language model for Apple Silicon — feed it images, get answers, all on-device. Abliterated Qwen3-VL 32B Instruct in 4-bit MLX. 19.6 GB, wants a 32 GB Mac. No cloud, no API key, no refusals.

A 4-bit MLX quantization of huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated, tuned for fast on-device vision-language inference on Apple Silicon.

• Base model: huihui-ai/Huihui-Qwen3-VL-32B-Instruct-abliterated (BF16 abliterated)
• Architecture: Qwen3-VL (dense 32B, multimodal image-text-to-text)
• Quantization: 4-bit affine, group size 64
• Format: MLX safetensors
• Footprint: ~17 GB on disk, comfortable on a 32 GB Mac and snappy on 64 GB+
• Context: 262 K tokens (256 K native, per upstream Qwen3-VL)
• Modalities: text + image input → text output

Why this exists

Qwen3-VL is currently the strongest open-weight vision-language model family in its size class. The 32B Instruct variant handles single-image, multi-image, and video inputs with the same 256 K-token context window the text-only Qwen3 series uses. huihui-ai's abliteration applies refusal-direction projection (Arditi et al., 2024) so the model doesn't refuse benign-but-edgy multimodal queries — important for security-research, content-moderation testing, OCR of unconventional documents, and any workflow where the upstream's safety scaffolding produces noise rather than safety.

As of publication, this is the first MLX-format quantization of any abliterated Qwen3-VL model. The non-abliterated MLX builds (lmstudio-community, mlx-community) cover 3-bit through 8-bit but none have abliteration. huihui's BF16 abliterated source pulls ~224,000 downloads / 30 days — there's a large audience that wants this on Apple Silicon.

What you can do with it

• Single-image visual Q&A on a 32 GB+ Mac, no cloud round-trip
• Multi-image reasoning (compare two diagrams, summarize a multi-page screenshot dump)
• Video understanding via per-frame ingestion (Qwen3-VL replaced T-RoPE with explicit textual timestamps)
• OCR-heavy tasks where cloud vision APIs would log the document content
• Document review for compliance-sensitive workflows (NDAs, medical records, sealed depositions) — see AirGap AI

Usage

Python (mlx-vlm)

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template

model, processor = load("divinetribe/Huihui-Qwen3-VL-32B-Instruct-abliterated-4bit-mlx")

messages = [
    {"role": "user", "content": [
        {"type": "image", "image": "https://example.com/your-image.jpg"},
        {"type": "text", "text": "What's in this image? Be specific."},
    ]},
]
prompt = apply_chat_template(processor, model.config, messages)
response = generate(model, processor, prompt=prompt, image=["https://example.com/your-image.jpg"], max_tokens=512)
print(response)

CLI

pip install mlx-vlm
python -m mlx_vlm.generate \
  --model divinetribe/Huihui-Qwen3-VL-32B-Instruct-abliterated-4bit-mlx \
  --image /path/to/image.jpg \
  --prompt "Describe what's in this image."

Local OpenAI-compatible server

python -m mlx_vlm.server \
  --model divinetribe/Huihui-Qwen3-VL-32B-Instruct-abliterated-4bit-mlx \
  --port 8080

That gives you a local http://localhost:8080/v1/chat/completions endpoint that accepts the standard OpenAI vision-message format. No API keys, no telemetry.

Where it fits in the lineup

Abliteration

"Abliteration" suppresses the model's built-in refusal direction so it doesn't refuse benign-but-edgy queries. It is not a general capability upgrade — use responsibly, and you remain bound by the upstream Qwen3-VL Apache-2.0 license.

Credits

• Qwen3-VL base by Alibaba Qwen team
• Abliteration by huihui-ai — the dominant abliteration shop on Hugging Face
• MLX 4-bit conversion by divinetribe using mlx-vlm 0.4.4

License

Apache 2.0, inherited from the upstream Qwen3-VL-32B-Instruct.

About the author

This model was built by Matt Macosko (@nicedreamzapp) for the claude-code-local stack — run Claude Code 100% on-device with local AI on Apple Silicon (⭐ 2,600+ on GitHub).

• 🤗 All my models: nicedreamzwholesale.com/software/huggingface/
• 💻 Software portfolio: nicedreamzwholesale.com/software/
• 🔒 AirGap AI (legal / healthcare / NDA workflows): nicedreamzwholesale.com/airgap/

More abliterated MLX models

Part of the Abliterated MLX for Apple Silicon
collection — 11 uncensored models converted for Macs, from Gemma 4 12B up to
Llama 3.3 70B, plus Qwen3, Qwen3-VL, Hermes 4 and Muse Glimmer 30B.

Part of Claude Code Local

This model is one of the fighters in Claude Code Local (3.2k★), which runs Claude Code 100% on-device on Apple Silicon through an MLX-native Anthropic-API server. Not sure which local model to run as an agent? Check the Agent-12 local agent leaderboard: real agent tasks, judged by the filesystem, same hardware for every row.

Built by Matt Macosko in Arcata, CA. Open to work on local-AI and Apple Silicon inference: [email protected].