microsoft/Fara1.5-4B
Fara1.5-4B is a 4 billion parameter multimodal computer use agent (CUA) developed by Microsoft Research AI Frontiers, based on Qwen3.5-4B. It operates on web browsers using vision-only perception from screenshots, emitting structured tool calls to complete end-to-end web tasks. With a 262K context length, it excels at automating repetitive web tasks like form filling, shopping, and booking, while incorporating critical-points safety for sensitive actions.
Loading preview...
Overview
Fara1.5-4B, developed by Microsoft Research AI Frontiers, is a 4 billion parameter multimodal computer use agent (CUA) designed for web browsers. Built upon Qwen3.5-4B, it operates by perceiving the browser solely through screenshots, without relying on the DOM or accessibility tree. It then predicts and executes actions like clicks, typing, scrolling, and navigation using structured tool calls to complete complex web tasks end-to-end.
Key Capabilities
- End-to-end web task completion: Automates tasks such as filling forms, making reservations, and managing shopping carts.
- Vision-only perception: Processes web pages exclusively from screenshots, mimicking human interaction.
- Coordinate-grounded actions: Predicts precise pixel-level targets for interactions, eliminating the need for separate grounding models.
- Critical-points safety design: Trained to pause and seek user confirmation before performing sensitive or irreversible actions like entering personal information, making payments, or submitting forms.
- Extended context length: Features a 262,144-token context window, supporting multi-screenshot trajectories and full action history.
- On-device viability: Its 4B parameter size allows it to run efficiently on single high-end GPUs.
Recommended Deployment
For secure and reliable operation, Fara1.5-4B is co-designed with and recommended to be deployed within MagenticLite. This harness provides essential features like sandboxing, allow-lists for navigation, real-time monitoring (watch-mode), and the ability to pause agent activity, ensuring responsible AI practices.
Evaluation
Fara1.5-4B demonstrates strong performance on end-to-end web agent benchmarks, achieving 80.8 on WebVoyager, 57.3 on Online-Mind2Web, and 27.4 on WebTailBench.