Microsoft has announced the launch of Fara-7B, a 7-billion-parameter language model designed specifically as a Computer Use Agent (CUA). It operates directly on users’ devices without relying on the cloud, enhancing privacy and reducing latency. This makes it ideal for organizations that need AI to handle critical tasks without sending data off-device, according to Microsoft Research on November 24, 2025.
Fara-7B functions like a human user, interpreting screenshots to click or type based on pixel information, without using browser accessibility trees meant for screen readers. This allows the model to handle complex or hidden web code more effectively. Its security approach, called pixel sovereignty by Yash Lara from Microsoft Research, ensures that all screen processing and reasoning remain on-device.
Tests on WebVoyager show Fara-7B scoring 73.5%, surpassing GPT-4o at 65.1% and UI-TARS-1.5-7B at 66.4%, using an average of only 16 steps compared to nearly 41 steps for competitors. This highlights its speed and efficiency advantages, with VentureBeat noting it offers the best cost-accuracy tradeoff among CUA models.
Despite its high performance, Microsoft warns of risks such as hallucinations or errors in complex tasks. To address this, researchers embedded Critical Points that require user approval before sending emails or executing irreversible transactions, preventing major mistakes.
Designing for safety without disrupting users remains a major challenge. Lara explained that interfaces like Magentic-UI allow precise user intervention, reducing approval fatigue from frequent confirmation prompts. This reflects the trend of highly capable AI agents even in smaller models.
 Origin: Venturebeat
