microsoft/Fara1.5-4B

Overview

Fara1.5-4B is a multimodal computer use agent (CUA) from Microsoft Research AI Frontiers, specifically engineered for web browsers. This 4.5 billion parameter model operates by perceiving the browser through screenshots, rather than the DOM, and then executing tasks by generating structured tool calls such as clicks, typing, and scrolling. It is supervised fine-tuned from Qwen3.5-4B using data synthesized by FaraGen1.5, a multi-agent pipeline.

Key Capabilities

• End-to-end web task completion: Automates complex tasks like form filling, reservations, job applications, and shopping carts.
• Vision-only perception: Processes screenshots to understand the browser state, mimicking human interaction.
• Coordinate-grounded actions: Predicts precise pixel-level targets for actions like clicks and drags.
• Critical-points safety design: Trained to pause and seek user confirmation before irreversible actions (e.g., payments, submissions) or when personal information is required.
• Extended context length: Features a 262,144-token context, enabling multi-screenshot trajectories and full action history.
• On-device viability: Its 4B parameter count allows it to run efficiently on single A100/H100/B200 GPUs.

Good For

• Automating repetitive web tasks such as filling forms, booking travel, and online shopping.
• Serving as a grounding model for other agents requiring pixel-accurate action prediction.
• Deployment within sandboxed environments like MagenticLite for enhanced security and control.

Limitations

• Primarily supports English; not suitable for other languages.
• Not recommended for high-stakes domains (legal, health, finance) due to potential for inaccuracies.
• Vision-only perception can be susceptible to deceptive page rendering or visual ambiguities.
• Multi-step trajectories can accumulate errors, and there is non-trivial run-to-run variance.