MagenticLite, MagenticBrain, Fara1.5: An Agentic Experience Optimized for Small Models
Key point
Microsoft has unveiled MagenticLite and Fara1.5, agentic systems designed for small models.
Details
Microsoft Research AI Frontiers has unveiled MagenticLite. This is the next generation of the existing Magentic-UI, an experimental agentic experience that handles the browser and the local file system in a single workflow.
The core is the combination of three components. MagenticBrain handles planning, delegation, and terminal use, while Fara1.5 is a family of computer-use models that handle browser-based tasks, designed to work together with it.
This release is based on the assumption that tool orchestration and action, rather than "more knowledge," determine agent performance. The result shows that even small models can perform a wide range of tasks while significantly lowering cost.
Fara1.5 is provided in three sizes—4B, 9B, and 27B—and is built on Qwen 3.5. In particular, the 9B model is recommended for most use cases, achieving SOTA among small computer-use models on Online-Mind2Web and nearly doubling web navigation performance compared to Fara-7B.
The improvements are as follows.
- Improved ability to handle real-world tasks such as form filling, logins, and reservations
- Added native action space and context retention for long-running tasks
- Designed so users can specify permissions or preferences at critical moments
- Enhanced data pipeline utilizing both realistic synthetic environments and the live web together
In addition, MagenticLite redesigned the browser and chat UI to better show users the agent's reasoning and actions, and to allow users to intervene directly when needed. In other words, it is a case of pushing forward the agentic experience for the era of small models by designing the model, tools, and harness to work together in an interlocking manner.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.