LethalDonkey/Qwen3-VL-8B-Instruct-abliterated-v2-MLX-4bit

Qwen3-VL-8B-Instruct-abliterated-v2 — MLX 4-bit

This is an MLX (4-bit) conversion of prithivMLmods/Qwen3-VL-8B-Instruct-abliterated-v2, an abliterated (refusal-removed) build of Qwen3-VL-8B-Instruct.

It runs natively on Apple Silicon (M-series) via Apple's MLX framework and mlx-vlm — typically faster than llama.cpp/Metal for image encoding, with no separate vision-encoder (mmproj) file needed.

As of conversion, no MLX build of this model existed — this is a community conversion to bring it to Apple Silicon users.

Use it in an app

This model is wired into the Qwen3-VL Captioner desktop app — pick it from the MLX section of the model dropdown on a Mac and it downloads + loads automatically.

Use it directly (mlx-vlm)

pip install mlx-vlm

from mlx_vlm import load, stream_generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model, processor = load("LethalDonkey/Qwen3-VL-8B-Instruct-abliterated-v2-MLX-4bit")
config = load_config("LethalDonkey/Qwen3-VL-8B-Instruct-abliterated-v2-MLX-4bit")

messages = [
    {"role": "system", "content": "You are a helpful assistant that describes images accurately and in detail."},
    {"role": "user", "content": "Describe this image in detail."},
]
prompt = apply_chat_template(processor, config, messages, num_images=1)

for chunk in stream_generate(model, processor, prompt, image=["your_image.jpg"], max_tokens=512):
    print(chunk.text, end="", flush=True)

Quantization

• Bits: 4
• Format: MLX (safetensors), converted with mlx_vlm.convert
• Choose 4-bit for the smallest size / lowest memory, 8-bit for the best quality, 6-bit for a balance.

Credits

• Base model: prithivMLmods/Qwen3-VL-8B-Instruct-abliterated-v2 by prithivMLmods
• Original architecture: Qwen3-VL by Qwen / Alibaba
• MLX framework: Apple (ml-explore/mlx)
• VLM tooling: mlx-vlm (Blaizzy/mlx-vlm)
• Converted for the Qwen3-VL Captioner project

License

Apache-2.0, inherited from the base model.