Microsoft Research Ships MagenticLite to Run Agents on Local Hardware

Fara1.5 nearly doubles the performance of its predecessor, Fara-7B, on web navigation tasks, marking a significant shift toward capable agents that run directly on consumer hardware.

Microsoft Research AI Frontiers has released MagenticLite, an experimental agentic application designed to bridge the gap between small model efficiency and complex task execution. Unlike traditional heavy-weight agents, MagenticLite operates across both the web browser and local file systems in a single, unified workflow. The system is built on a core research hypothesis: that true agentic capability stems from sophisticated tool orchestration and action rather than sheer parameter count.

Orchestration via MagenticBrain and Fara1.5

The system relies on two purpose-built models working in tandem. MagenticBrain, a 14B-parameter model fine-tuned from Qwen 3, acts as the central orchestrator. It handles reasoning, planning, and code generation, turning vague natural-language requests into actionable steps. Crucially, MagenticBrain is trained to recognize when a task requires specialized browser interaction and issues a structured handoff to the computer-use subagent.

That subagent is Fara1.5, a family of models (4B, 9B, and 27B) based on Qwen 3.5. The flagship 9B model sets new state-of-the-art results on the Online-Mind2Web benchmark, outperforming similarly sized models and achieving over 90% performance in its largest 27B variant. Fara1.5 is specifically tuned for "computer-use," meaning it excels at navigating credentialed sites, filling out complex forms, and managing long-running tasks that span hundreds of steps.

A Harness Built for Small Model Constraints

Running agents on smaller models requires overcoming significant technical hurdles, particularly regarding context window degradation. The MagenticLite harness addresses this through active context management. Instead of feeding the entire history into the model, the harness curates prompts at each step, condensing earlier interactions into concise summaries and surfacing only the most relevant information to keep the models focused.

To ensure safety and reliability, the system incorporates human-in-the-loop guarantees. It uses "critical point" detection to pause execution for explicit user approval during sensitive actions, such as financial transactions or irreversible file deletions. Furthermore, the entire execution environment is isolated within Quicksand, an open-source wrapper providing a QEMU-based sandbox that protects the host system from browser sessions and generated code.

This release signals a broader move toward decentralized AI, where capable, specialized agents perform meaningful work locally without the latency or privacy concerns of massive cloud-based models.

Source: MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models
Domain: microsoft.com