Microsoft unveiled Fara-7B earlier this week, an experimental agentic small language model built to operate computers directly, 7 billion parameters, designed to run on-device for lower latency and improved privacy.
Fara-7B is Microsoft Research’s first Computer Use Agent small language model (SLM) that interacts with software the way a human does: by seeing the screen and issuing mouse and keyboard actions rather than only producing text. Unlike chat-only models, it predicts coordinates for clicks, types into fields, and issues browser macro-actions, enabling it to complete multi-step web tasks end-to-end on behalf of users.
The team trained Fara-7B using a novel synthetic data pipeline that generates multi-step web task demonstrations from public websites and task seeds. The final training set included 145,000 trajectories totaling 1 million steps, distilled into a single 7B model so it can run natively on devices without orchestrating multiple large models. The model uses Qwen2.5-VL-7B as its multimodal base and was trained with supervised fine-tuning rather than reinforcement learning.
Fara-7B achieves state-of-the-art results for its size, outperforming comparable 7B agents and competing with much larger systems on benchmarks such as WebVoyager and a new Microsoft benchmark called WebTailBench. For example, Fara-7B posts strong task success rates and sits on a favorable cost-performance frontier compared to agents that prompt larger LLMs. Microsoft is releasing Fara-7B as open-weight on Microsoft Foundry and Hugging Face, with a quantized, silicon-optimized build available for Copilot+ PCs running Windows 11 for on-device experimentation.
Because agents that operate computers can take real-world actions, Microsoft built safety into Fara-7B from the ground up. The model logs every action for auditability, is intended to run in sandboxed environments, and enforces Critical Points, places where it must pause and request user consent before proceeding with transactions or personal-data actions. In red-team style evaluations, Fara-7B showed a high refusal rate on harmful or risky tasks (82% on WebTailBench-Refusals) and underwent Microsoft’s internal red teaming and safety checks.
Fara-7B is an AI shift toward practical, private, on-device agentic assistants that can automate routine web tasks like shopping, booking, and information retrieval without sending full context off-device. By open-sourcing the model and tooling (including Magentic-UI integrations), Microsoft aims to accelerate community experimentation while emphasizing responsible deployment and ongoing research into accuracy and safety.
If you want to try Fara-7B, Microsoft provides inference code and pre-optimized builds for local testing, and encourages sandboxed exploration and feedback as the technology matures.
