microsoft/fara

What it solves

Fara1.5 addresses the difficulty of creating AI agents that can actually use a computer like a human. Instead of relying on text-based accessibility trees or separate parsing models, it enables agents to perceive a browser visually and execute multi-step tasks using mouse and keyboard inputs.

How it works

Fara1.5 is a family of native Computer Use Agents (CUAs) available in three sizes (4B, 9B, and 27B) built on Qwen3.5. It operates in an "observe-think-act" loop: the model takes a screenshot of the browser and the conversation history, reasons about the current state, and then predicts specific coordinates for mouse clicks or keyboard inputs.

To train these models, Microsoft developed FaraGen1.5, a scalable data pipeline that uses synthetic environments (clones of mail, calendars, etc.) and a system of solvers and verifiers to generate high-quality training data without relying solely on expensive human demonstrations.

Who it’s for

It is designed for researchers and developers building autonomous web agents, as well as users looking to automate everyday web tasks like booking travel, comparing prices, or filling out forms.

Highlights

• Visual Operation: Interacts with webpages via direct coordinate prediction rather than accessibility trees.
• Compact Scale: Available in 4B, 9B, and 27B parameter versions for lower latency and better privacy.
• Human-like Interaction: Trained to ask for missing information, flag ambiguity, and pause for approval before taking irreversible actions.
• Comprehensive Benchmarking: Includes WebTailBench for real-world tasks and CUAVerifierBench for evaluating the judges that score agent performance.