In a move that could bring advanced automation directly to personal devices, Microsoft Research has unveiled Fara-7B, a new type of artificial intelligence model built specifically to operate a computer. Unlike conventional chatbots, this compact model is designed to visually perceive a screen and perform tasks—like clicking, typing, and scrolling—on a user's behalf.

The 7-billion-parameter model is notable for its small size, which allows it to run efficiently on local devices such as laptops. This enables faster response times and keeps user data private, as information does not need to be sent to the cloud for processing.

How Fara-7B Works

Unlike complex systems that chain multiple large models together, Fara-7B is a single, efficient model. It interacts with a computer much like a human would:

1. It sees the screen, processing webpage screenshots without relying on behind-the-scenes code or accessibility data.
2. It reasons about the task, deciding the next logical step.
3. It takes direct action, outputting commands to click, type, scroll, or search.
4. This streamlined approach allows it to run directly on a device, such as a Copilot+ PC, which keeps user data local and reduces response latency.

Designed for Real-World Use

To train Fara-7B, Microsoft developed a novel method to generate synthetic training data at scale. The system used a multi-agent framework to practice completing thousands of diverse tasks—from comparing product prices to booking movie tickets—on real, public websites. Only successful task trajectories were used to train the final model.

In evaluations, Fara-7B demonstrated strong performance against larger, more resource-intensive systems. It achieved state-of-the-art results for its size class on established benchmarks and a new one introduced by Microsoft, WebTailBench, which focuses on practical tasks like making reservations and applying for jobs.

Safety, Availability, and Next Step

Because computer-use agents perform real actions, safety and user control are paramount. Fara-7B is designed to stop and ask for user consent at "Critical Points," such as before entering personal data or completing a purchase. Microsoft recommends running the model in a sandboxed environment as it remains an experimental research preview.

Fara-7B is being released under an open-weight MIT license. It is available now on Microsoft Foundry and Hugging Face, and a version optimized for Copilot+ PCs will be accessible through the AI Toolkit in Visual Studio Code.

This release aims to accelerate community exploration into how small, efficient models can become helpful, private, and integrated agents in our daily digital workflows.