Microsoft's MagenticLite brings capable AI agents to small models

Microsoft Research AI Frontiers has released MagenticLite, the next-generation agentic application designed specifically for small models. It combines a redesigned agent harness with two purpose-built models: MagenticBrain, which handles reasoning, planning, coding, and delegation, and Fara1.5, a computer-use model family optimized for browser tasks. The flagship Fara1.5 model has 9 billion parameters and sets a new state-of-the-art among small computer-use models, nearly doubling the web navigation performance of its predecessor Fara-7B. MagenticLite works seamlessly across both the browser and local file system, enabling a broad range of agentic tasks while keeping all data on the user's hardware.

The project embodies a research bet that agentic capability depends more on tool orchestration and action than raw knowledge, making it feasible to use smaller models efficiently. The team redesigned the full lifecycle—data generation, training objectives, model design, and orchestration—to achieve reliable performance at this scale. User experience improvements include visibility into the agent's reasoning, the ability to take direct control, and explicit approval at critical points. MagenticLite is built for real-world tasks like filling forms, browser research, and local file management, with evaluations grounded in scenario-based testing alongside standard benchmarks.