In a significant move for the artificial intelligence sector, Microsoft has launched Fara-7B, its first small language model (SLM) designed to operate a computer like a human. Announced on November 28, 2025, this compact 7-billion-parameter model marks a shift from bulky, cloud-dependent AI systems to efficient, on-device intelligence.
What Makes Fara-7B a Game-Changer?
Fara-7B is not a typical text-generating AI. It is a Computer Use Agent (CUA) built to perform actions such as clicking, typing, and navigating by simply analyzing a screenshot of a computer screen. Unlike traditional CUA models that rely on massive cloud servers and multiple subsystems, Fara-7B is a single, self-contained model that can run directly on a user's device.
Microsoft has described it as its first agentic SLM for computer use, highlighting its surprising power despite its small size. This local operation results in reduced latency and enhanced privacy, as sensitive user data does not need to be sent to the cloud.
How Was Fara-7B Developed and Trained?
The creation of Fara-7B was powered by a massive synthetic data pipeline called FaraGen. This system trained AI agents on real websites across an impressive 70,000 domains. It generated realistic multi-step sessions that mimicked human behavior, complete with actions like scrolling, searching, and even making mistakes.
To ensure quality, every session was reviewed by three separate AI judges. After a rigorous filtering process, Microsoft used 145,630 verified sessions containing over 1 million individual actions to train the final Fara-7B model.
Performance, Cost, and Availability
When it comes to performance benchmarks, Fara-7B holds its own. It achieved 73.5% on Web Voyager, 34.1% on OnlineMind 2 Webb, 26.2% on DeepShop, and 38.4% on WebTailBench. The WebTailBench result is particularly crucial as it tests real-world tasks like job applications and property searches.
The economic advantage is stark. Microsoft estimates that a full task with Fara-7B costs around 2.5 cents, a fraction of the roughly 30 cents required by large-scale agents using models like GPT-4.
In a push for wider developer adoption, Fara-7B is now available on Microsoft Foundry and Hugging Face under an MIT license. It is integrated with Magentic-UI, a Microsoft Research prototype. Furthermore, a quantised, silicon-optimised version for Copilot+ PCs on Windows 11 is being released, allowing users to download and test the model locally on their devices.
By making this model open-weight, Microsoft aims to lower the barrier for developers and accelerate innovation in automating everyday web tasks, paving the way for a more accessible and private AI future.
