MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models

Details

Microsoft Research AI Frontiers has released three components designed around a single research thesis: agentic capability depends on tool orchestration and action, not raw model size or knowledge. This bet enables capable agents from small models at a fraction of the cost of frontier API calls. MagenticLite is the next generation of Magentic-UI, rebuilt with a harness optimized for small models and extended to work across both the browser and local file system in a unified workflow. Data stays on the user's machine. MagenticBrain is the system's planner and orchestrator: it decomposes vague requests into concrete plans, selects tools and subagents, writes code when needed, and recovers from mid-task failures. Fara1.5 is the computer-use model family for browser tasks, available in three sizes with a 9-billion-parameter flagship that Microsoft claims 'nearly doubles Fara-7B's performance on web navigation' and sets new state-of-the-art results among small computer-use models, with improved handling of forms, credentialed sites, and long-horizon tasks. The three components are codesigned to work together but are each useful independently. The longer-term goal is explicit: capable AI agents running directly on user hardware without cloud dependency — a privacy-preserving architecture that would meaningfully expand where agents can operate.