Microsoft's tiny AI agent can click and type for you
Microsoft just open-sourced Fara-7B, a compact 7B-parameter model specifically built to take over keyboard and mouse actions. By reading raw browser screenshots instead of code trees, it automates routine web tasks like booking trips or comparing prices. Why it matters: It is small enough to run locally on Copilot+ PCs, delivering fast agentic workflows without uploading private screen data.
Fara-7B is a 7-billion-parameter open-weight model from Microsoft built specifically to click, type, and complete browser tasks on your behalf.
Automating routine web tasks usually meant writing fragile scraper scripts or paying heavy API bills for multi-agent frontier systems. Fara-7B takes the opposite approach by distilling full browser-use capabilities into a compact model that runs directly on your hardware.
What it does
Fara-7B interacts with a web browser just like a human does. It does not parse accessibility trees or read raw DOM code. Instead, it looks directly at raw screenshots of the browser window.
At every step, the model ingests your original prompt, the past action history, and the last three screenshots. It outputs a brief thought trace before issuing a specific tool call, such as clicking exact screen coordinates, typing into fields, scrolling, or triggering navigation commands like searching and opening URLs.
It handles everyday multi-step workflows like comparing prices across multiple online retailers, booking event or restaurant tickets, searching job boards, and filling out complex online forms.
Under the hood and benchmark results
Microsoft built Fara-7B on top of Qwen2.5-VL-7B, capitalizing on its strong visual grounding and 128k context support. It was trained entirely with supervised fine-tuning, skipping reinforcement learning altogether.
To overcome the lack of high-quality human demonstration data, the team created an automated multi-agent synthetic pipeline using Magentic-One. Teams of agents planned tasks, executed multi-step trajectories across live web pages, and verified results with multi-modal checks, generating a curated dataset of 145,000 trajectories spanning 1 million action steps.
On the standard WebVoyager benchmark, Fara-7B scored a 73.5% success rate under LLM-as-judge evaluation, and hit 38.4% on Microsoft's newly released WebTailBench, outperforming larger setups like OpenAI's computer-use-preview. In an independent human-verified test hosted by Browserbase, it secured a 62% success rate with retries enabled.
Because this is an experimental 7B model, it still has rough edges. It can make mistakes on long, winding tasks, hallucinate elements, or fail complex prompt instructions. Microsoft explicitly advises running it inside a sandbox and keeping high-risk credentials or sensitive personal data out of reach.
How to get started
Fara-7B is fully open under the permissive MIT license, available today on Hugging Face and Microsoft Foundry.
If you own a Windows 11 Copilot+ PC, you can download a pre-quantized, silicon-optimized version designed for turnkey local execution. The release also integrates seamlessly with Microsoft's Magentic-UI research prototype for quick local experimentation.
Why it matters
If you build local agents or automation tools, Fara-7B is one of the most practical releases in months. It proves that a tiny 7B vision-language model, trained on carefully vetted synthetic trajectories, can match or beat cumbersome multi-model agent stacks.
More importantly, running a computer-use agent locally eliminates both runaway cloud token costs and massive privacy risks. Screen captures never leave your machine, making local execution genuinely viable for personal desktop tasks, provided you keep a human in the loop to supervise tricky actions.
Sources (1)
- microsoft.com(opens in a new tab)Originalmicrosoft.com