Recommended MLX 4-bit, group-size 64 conversion of Ocelot-1-VL, merged into its BF16 Qwen3-VL-4B-Instruct base. Effective quantization is 5.577 bits/weight because sensitive and unsupported tensors remain at higher precision.

This model is specialized only for webpage summarization. Follow the strict prompt contract and limitations in the original model card.

bash

pip install 'mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git'
python -m mlx_vlm generate --model . --prompt 'The is the text of a webpage: <page>Page text here</page> Summarise the content between the <page> tags, or if no content is found use the screenshots provided, in the Brave Summary style.' --max-tokens 512

For screenshots, add --image webpage.png and begin the prompt with The following is a screenshot of a webpage:. Converted with MLX-VLM revision 0b1d25e334686bd36dda71b2307d186dbb3e7859. Text and screenshot tests passed. An Apple M4 Pro test used 3.34 GB peak memory and measured 45 prompt tokens/s and 15 generation tokens/s.

gnukeith/Ocelot-1-VL-MLX-4bit

Ocelot-1-VL MLX 4-bit

gnukeith/
Ocelot-1-VL-MLX-4bit