Microsoft released Fara-7B, a 7-billion-parameter AI agent designed to run locally on a user's PC to automate web-based tasks.
The model operates by visually perceiving screenshots to control the mouse and keyboard, an approach intended to enhance user privacy by keeping data off the cloud.
While competitive with larger models like GPT-4o, Microsoft is releasing Fara-7B as an experimental model, warning that it can still make mistakes.
Microsoft has released Fara-7B, a 7-billion-parameter agentic AI designed to operate directly from a user's PC. The model automates web tasks by controlling the mouse and keyboard, with its local-first approach intended to boost user privacy by keeping data off the cloud.
Taught to see: Unlike assistants that need to read a website's code, Fara-7B operates by visually perceiving screenshots, allowing it to navigate sites like a human. To train it, Microsoft built a synthetic data engine called FaraGen, which, according to an accompanying research paper, generates millions of verified action steps across tens of thousands of live websites.
Punches above its weight: The result is a compact model that reportedly holds its own against bigger systems. On industry benchmarks, Fara-7B is competitive with larger models like GPT-4o and tops others in its own weight class, all while completing tasks at a fraction of the cost.
Handle with care: But don't fire your assistant just yet. Microsoft is upfront that Fara-7B can still make mistakes or "hallucinate," and demos showed it working slower than a human. The company is releasing it as an experimental, open-weight model and advises users to stick to sandboxed environments for now.
Fara-7B marks a concrete step toward capable, on-device AI agents that can automate daily digital chores. However, its acknowledged limitations show that the path to a fully autonomous, reliable AI assistant is still a long one. The model is being integrated with Magentic-UI, a research prototype from Microsoft's AI Frontiers group where developers can test its capabilities. It builds on the company's broader push into on-device AI, following its release of previous small models like Phi.
