In the field of artificial intelligence, Microsoft Research AI Frontiers Lab recently released the new Fara1.5 series of agent models. This series is specifically designed for browser scenarios, aiming to enhance the level of intelligence in computer usage. The Fara1.5 series includes three versions with different parameter scales: 4B, 9B, and 27B.

This agent model works with the MagneticLite sandbox browser interface, which can directly read screenshots from the browser and automatically complete various web tasks by simulating mouse and keyboard operations. The workflow of Fara1.5 is based on the "Observe - Think - Act" cycle, where each step combines historical conversations and the latest three browser screenshots to generate reasoning content and subsequent actions.

image.png

In terms of performance, Fara1.5-27B achieved a task success rate of 72% in the Online-Mind2Web benchmark test, significantly higher than competitors such as OpenAI's Operator (58.3%) and Gemini2.5Computer Use (57.3%). Even the Fara1.5-9B version reached a success rate of 63.4%. This efficient performance has made the Fara1.5 series stand out among many agents.

The training of Fara1.5 used approximately 2 million samples for fine-tuning, with 60% coming from web trajectories, 12.8% from synthetic environments, 12.5% related to user interactions, 8.8% as factual anchors, and 4.9% from visual question answering. This diverse and rich training data greatly enhanced the model's adaptability and intelligence level.

To ensure user safety and privacy, Fara1.5 will actively stop operations and ask users under the following three circumstances: lack of personal information, unclear task descriptions, and impending execution of unapproved irreversible operations. In addition, all operation records will be stored in the MagneticLite sandbox, ensuring a secure boundary between the agent and the user's device.

With continuous technological advancements, Microsoft's Fara1.5 series undoubtedly brings users a more intelligent and convenient browsing experience, demonstrating the huge potential of artificial intelligence in daily computer use.

Key points:

🌟 The Fara1.5 series model has a task success rate of up to 72%, surpassing competitors like OpenAI.

🖥️ This model completes web tasks through a sandbox browser interface, with an intuitive and easy-to-use operation method.

🔒 Emphasizes user safety, proactively asks about ambiguous tasks and personal information protection to ensure privacy security.