---
language:
- en
library_name: other
tags:
- browser-agent
- ui-agent
- dom-agent
- safe-readonly
license: mit
---

# LEO-UI7M v0.1

Lightweight DOM-based browser agent. 7.4M parameters, CPU-only, no screenshots required.  
Reads page DOM structure, predicts the next UI action (click, type, navigate).

## Quick Start

```bash
pip install -r requirements.txt
python examples/run_agent.py --task "find the price of item X"
```

## Model

- Architecture: DOM-aware action predictor
- Parameters: 7,458,567
- Weights: model.safetensors
- GPU required: No
- Input: raw DOM text + task instruction
- Output: UI action (element + action type)

## Validated Capabilities

- Wikipedia command → research note
- Python docs command → structured docs results
- Hacker News jobs command → job digest
- Page explanation from URL
- Refusal for unsafe or out-of-scope tasks

## Limitations

- Not a general browser automation replacement
- Not screenshot-based (DOM text only)
- Not for arbitrary web automation
- Best on pages with clean, semantic DOM structure
- Trained on limited data — generalizes to simple, well-structured pages

## Files

```
model.safetensors     — model weights
config.json           — model config
leo_ui7m/             — Python package (runtime, adapters, agent)
examples/run_agent.py — usage example
requirements.txt      — dependencies
tests/                — test suite
```

## License

MIT