Prime Highlight
- Microsoft has launched Fara-7B, a 7B-parameter Computer Use Agent designed to execute tasks directly on a user’s device through mouse, keyboard, and visual interaction.
- Despite its small size, Fara-7B delivers performance comparable to larger models and can automate tasks like web navigation, summarization, and online purchases.
Key Facts
- Fara-7B achieved a 73.5% success rate on the WebVoyager benchmark, outperforming GPT-4o’s score of 65.1%.
- The model is available via Microsoft Foundry and Hugging Face under the MIT license, with an optimized version planned for Windows 11 Copilot+ PCs.
Background
Microsoft has introduced Fara-7B, a new small language model built to perform computer tasks directly on a user’s device using standard interfaces such as the mouse and keyboard. Described as a Computer Use Agent (CUA), the 7-billion-parameter model allows AI to visually read web pages and carry out actions including scrolling, typing, clicking, and navigating online platforms.
Announcing the launch, Microsoft said Fara-7B enables developers to create agentic AI experiences that automate everyday digital tasks. The company demonstrated the model completing online purchases, summarizing webpages, and calculating driving time using online maps. Despite its small size compared to traditional large language models, the company claims Fara-7B delivers performance on par with much larger systems.
On the WebVoyager test, which checks how well an AI agent completes web tasks, Fara-7B finished 73.5% of tasks successfully. It did better than GPT-4o, which scored 65.1%. Microsoft noted that the model’s compact architecture also reduces latency and boosts privacy, as tasks can be processed locally without sending data to the cloud.
The model was trained on synthetic datasets generated from real web pages and real user tasks. However, Microsoft emphasised that Fara-7B remains experimental. Current limitations include reduced accuracy on complex workflows, occasional failures to follow instructions, and susceptibility to hallucinations. The company said it is actively collecting user feedback to refine the system.
Fara-7B is now available through Microsoft Foundry and Hugging Face under the MIT license. Developers can test it using Magentic-UI, Microsoft’s AI research interface. The company also confirmed plans to release an optimized version for Windows 11 Copilot+ PCs, which have built-in hardware designed to run AI models efficiently.
