This week, Microsoft launched Fara-7B, its first agentic small language model designed specifically for computers.
Microsoft said the model uses computer interfaces, such as a mouse and keyboard to “complete tasks on behalf of users,” according to a blog post announcing the launch.
This means the model can visually see and interpret a webpage and perform actions such as scrolling, typing and clicking.
“Users can build and test agentic experiences beyond pure research — automating everyday web tasks like filling out forms, searching for information, booking travel, or managing accounts,” Microsoft wrote.
In three videos on its blog, Microsoft showed Fara-7B completing online purchases, summarizing information from web pages and using online maps to identify the driving time between two given locations.
Spanning 7 billion parameters, the model is far smaller than today’s most powerful large language models. For instance, when GPT-3 debuted in 2020, it featured more than 175 billion parameters.
Yet Microsoft said its latest model is still powerful enough to rival far larger systems. When tested as a computer-use agent on the WebVoyager benchmark for task success rate, it scored a 73.5%, surpassing GPT-4o , which scored a 65.1%.
In addition, Microsoft said its smaller size improves latency and privacy, as user data remains contained locally.
Fara-7B was trained on synthetic data sourced from real web pages as well as tasks from human users.
Microsoft stressed that Fara-7B remains in the experimental phase, with the company collecting user feedback to improve future iterations. In particular, the company highlighted challenges in accuracy on more complex tasks, mistakes following instructions and susceptibility to hallucinations.
“These are active areas of research, and we’re committed to ongoing improvements as we learn from real-world use,” Microsoft wrote.
Fara-7B is now available on Microsoft Foundry and Hugging Face under an MIT license. Developers can test the model through Magentic-UI, the company’s prototype AI research platform.
Looking ahead, the company said it will release another iteration optimized for Windows 11 Copilot+ PCs, which include dedicated hardware for running AI models.