Microsoft Research announced Fara-7B, an open-weight agentic small language model (SLM) built specifically for computer use. Instead of just generating text, Fara-7B operates a browser like a human—seeing screenshots, moving the mouse, typing, scrolling, and using tools to complete tasks. Official blog: Fara-7B: An Efficient Agentic Model for Computer Use.

• Purpose-built for computer use: Fara-7B is a 7B-parameter model that visually perceives webpages and executes actions (click, type, scroll, visit_url, web_search) without relying on accessibility trees.
• On-device capable: The compact size lets it run directly on devices, including a quantized, silicon-optimized version for Copilot+ PCs, improving latency and privacy since data can stay local.
• Training pipeline: Microsoft created a synthetic data generation system (building on AgentInstruct and Magentic-One) that produced 145K trajectories and 1M steps from real websites covering tasks like shopping, booking travel, job search, and price comparison.
• Benchmark performance: On benchmarks such as WebVoyager, Online-Mind2Web, DeepShop, and the new WebTailBench, Fara-7B reaches state-of-the-art task success rates versus other computer-use agents and even larger LLM-based systems, while using far fewer steps and tokens.
• Availability: The model is released as open weights under an MIT license on Microsoft Foundry and Hugging Face, enabling developers to experiment with real-world web automation in sandboxed environments.

Why it matters: Fara-7B shows that small, efficient models can power capable computer-use agents, bringing automated web actions—like filling forms, booking travel, or managing accounts—closer to running privately on personal devices. It marks a key step in making agentic computing practical beyond large, cloud-only LLM stacks.