Microsoft today unveiled its vision for the future of everyday consumer AI agents by introducing its new Fara-7B model—a lightweight AI capable of running locally on personal computers and intelligent enough to handle common computing tasks.
In a blog post, the company described Fara-7B as its first small language model specifically engineered for “computer use” tasks, with the ability to control a mouse and keyboard directly.
With just 7 billion parameters, Fara-7B is significantly smaller than today’s most powerful large language models. For context, even GPT-3—released back in 2020 before the current AI boom—boasted over 175 billion parameters. Despite its compact size, Microsoft claims Fara-7B delivers “state-of-the-art performance,” particularly when benchmarked against other models of similar scale.
The company further asserts that Fara-7B remains competitive even against larger, more resource-intensive agent systems that rely on multiple large language models. For instance, when fine-tuned specifically for web browsing, it reportedly outperforms OpenAI’s GPT-4o.
Unlike other systems that depend on auxiliary data structures like accessibility trees or separate vision models to interpret screen content, Fara-7B interacts with websites visually—just as a human would—using only what’s displayed on the screen.
Microsoft demonstrated Fara-7B’s capabilities in three videos, showing it complete tasks such as purchasing products online, researching information and summarizing findings, and measuring distances between locations using online maps—all initiated through simple user prompts. While the model executes these actions noticeably slower than a human and requires user approval at critical steps (like entering login credentials), the demos offer a compelling glimpse into a future where AI routinely automates daily digital chores—especially as models become faster and more capable.
While Microsoft’s existing Copilot can also act as an agent to automate tasks on behalf of users, it differs fundamentally in that it cannot run locally. Copilot relies entirely on Microsoft’s cloud-based infrastructure, requiring a constant internet connection and transmitting substantial user data from the PC—raising potential privacy concerns, despite the company’s safeguards against misuse of sensitive information.
In contrast, Fara-7B runs entirely on-device, installed directly on the user’s PC and leveraging only local hardware resources. This design eliminates data transmission to the cloud, reducing latency and enhancing privacy, according to Microsoft. The model builds on the company’s prior work with compact AI, including last year’s Phi-4, which was small enough to operate on smartphones.
Nevertheless, Microsoft acknowledges that Fara-7B isn’t flawless. During testing, it occasionally struggled with complex tasks, exhibited inaccuracies, and sometimes misinterpreted instructions. Like many generative models, it remains somewhat prone to hallucinations.
Due to these reliability limitations, Microsoft is initially releasing Fara-7B only within an isolated sandbox environment, where its behavior can be monitored and users are prevented from submitting sensitive information. The company has also implemented safety mechanisms to ensure the model rejects any potentially malicious prompts.
Fara-7B will be made available under the MIT license on Microsoft Foundry and Hugging Face—but exclusively for use with Magnetic-UI, Microsoft’s experimental AI research platform. Looking ahead, the company plans to release a specialized version optimized for Windows 11 Copilot+ PCs, which feature dedicated AI acceleration hardware.
