Seven-billion-parameter system can control a user’s mouse and keyboard, run locally for privacy, and outperform larger AI models in automated computer-task benchmarks.
Microsoft has quietly introduced Fara-7B, a new “Agentic” small language model designed to run directly on personal computers and carry out digital tasks by taking control of the mouse and keyboard. Unlike traditional AI assistants focused on answering prompts, Fara-7B acts as a computer-use agent capable of navigating websites, clicking, typing and scrolling — all by visually interpreting the screen in front of it.
The model, which contains just seven billion parameters, relies on the same sensory cues a human user would. According to Microsoft, Fara-7B perceives webpages visually and predicts precise on-screen coordinates to take action, eliminating the need for external parsing tools or accessibility-tree data. Its compact size allows it to operate locally on devices, reducing latency and ensuring user data stays private.
Microsoft says the project is a response to growing interest in small language models (SLMs), which are emerging as efficient alternatives to large language models for targeted tasks. Running Fara-7B entirely on-device avoids cloud processing, making it more cost-effective, faster and more secure.
Read Also: Bill Gates Launches $1.4B Program To Transform African Farming
Developing the model required large volumes of computer-interaction data, something researchers identified as a major obstacle. Instead of relying on human annotators — a process Microsoft says is prohibitively expensive — the team built a synthetic data pipeline. It generated multi-step tasks modeled on real web interactions, trained the system to attempt them, and fine-tuned its responses by filtering out failed attempts.
In total, Fara-7B was trained on 145,000 “trajectories” covering one million individual task steps across diverse websites and difficulty levels. Additional datasets helped reinforce skills such as UI-element detection, captioning and visual question-answering. In benchmark tests, Microsoft reports that Fara-7B often outperformed much larger models, including GPT-4o, in computer-use automation tasks.
The company stresses that Fara-7B remains experimental. It is designed for sandboxed environments where users can observe its behavior and prevent exposure to sensitive data. Built-in safeguards based on Microsoft’s Responsible AI Policy include an 82% refusal rate for inappropriate tasks and automatic halting whenever user consent or personal information is required.
Developers and researchers can now access Fara-7B through Microsoft Foundry, Hugging Face and the Magnetic-UI research platform. Microsoft is also releasing an optimized, quantized version intended to run on Copilot+ PCs powered by Windows 11.
By making the model open-weight, Microsoft says it hopes to lower barriers for the AI community to experiment with, adapt and improve computer-use automation — potentially transforming everyday tasks such as booking travel, managing accounts or completing online forms.