README.md

---
license: apache-2.0
base_model: [bravesoftware/Ocelot-1-VL, Qwen/Qwen3-VL-4B-Instruct]
library_name: mlx
pipeline_tag: image-text-to-text
tags: [mlx, qwen3-vl, vision, summarization, 4-bit]
---

# Ocelot-1-VL MLX 4-bit

Recommended MLX 4-bit, group-size 64 conversion of [Ocelot-1-VL](https://huggingface.co/bravesoftware/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.

These are final MLX weights, not conversion inputs. Users can open a local browser interface directly after installing the MLX runtime:

```bash
pip install 'mlx-vlm @ git+https://github.com/Blaizzy/mlx-vlm.git'
mlx_vlm.chat_ui --model gnukeith/Ocelot-MLX
```

The runtime downloads the model from Hugging Face automatically. No cloning, conversion, or Python code is required.

Direct command-line inference is also available:

```bash
mlx_vlm.generate --model gnukeith/Ocelot-MLX --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.