Agentic AI
Microsoft Research Releases MagenticLite: An Agentic AI System Optimized for Small Models
Microsoft AI Frontiers debuts MagenticLite, a new agentic application powered by purpose-built small models that deliver capable agent performance while running efficiently on user hardware.
Microsoft Research Blog
Microsoft Research AI Frontiers has released MagenticLite, an experimental agentic application designed to work across browsers and local file systems in unified workflows. The platform represents a significant shift in how capable agents can run directly on user hardware without requiring massive model sizes.
The Three-Component System#
MagenticLite combines three newly released components, each designed to work together:
- MagenticLite: The next-generation interface and harness, rebuilt from Magentic-UI with a focus on small-model execution
- MagenticBrain: A 14-billion-parameter orchestration model that plans, codes, and delegates tasks
- Fara1.5: A computer-use model family (4B, 9B, and 27B variants) for browser-based tasks
Performance Gains with Smaller Models#
The core research insight driving the project is that "agentic capability depends on tool orchestration and action rather than knowledge alone." This enables smaller models to handle complex tasks at a fraction of the cost.
Fara1.5 achieves state-of-the-art results among small computer-use models, nearly doubling the performance of its predecessor Fara-7B on web navigation tasks. The flagship 9B model improves handling of forms, credentialed sites, and long-running tasks. On the Online-Mind2Web benchmark—which contains 300 tasks across widely used web domains—Fara1.5 sets new performance records for its size class, with the larger 27B variant achieving over 90% performance.
Real-World Use Cases#
The system handles a broad range of agentic tasks, including:
- Filling out forms and making appointments
- Browser research and information analysis
- Managing local files
- Handling logins for credentialed sites
Design and Human Control#
The harness preserves human-in-the-loop controls, requiring explicit user approval at critical points. All operations run inside Quicksand, an open-source QEMU-based sandbox, isolating browser sessions and code execution from the host system.
