Microsoft Unveils Fara-7B, an Open-Weight Agentic Model for Computer Use

Microsoft announced in a research blog the release of Fara-7B, a 7 billion parameter small language model (SLM) designed as an agentic Computer Use Agent (CUA). Unlike text-based chat models, Fara-7B can perform actions such as clicking, typing, and scrolling to complete computer tasks directly through the user interface.

The model operates by visually perceiving webpages and executing predicted actions without relying on accessibility trees or additional parsing systems. Built on Qwen2.5-VL-7B, Fara-7B achieves state-of-the-art performance among models of its size and competes with larger, more resource-intensive systems. It can run directly on devices, improving speed and maintaining data privacy.

Fara-7B was trained using a synthetic data pipeline that generated 145,000 browser interaction trajectories covering diverse web tasks. These include activities such as booking travel, comparing prices, and applying for jobs. The model achieved 73.5% task success on the WebVoyager benchmark and 38.4% on the new WebTailBench, outperforming other 7B-class agents like UI-TARS-1.5-7B.

The model is available on Microsoft Foundry and Hugging Face under an MIT license. A quantized version optimized for Copilot+ PCs powered by Windows 11 can be downloaded through the AI Toolkit in Visual Studio Code for on-device experimentation. Fara-7B includes built-in safety mechanisms, such as halting at critical points requiring user consent, and is recommended for use in sandboxed environments.

By releasing Fara-7B as an open-weight model, Microsoft aims to support research and community experimentation in developing efficient computer-use agents capable of automating real-world web tasks.