MagenticBrain-14B

An 14-B orchestrator model for MagenticLiteMagenticBrain is an orchestration model developed by Microsoft Research AI Frontiers. It plans and coordinates multi-step agentic tasks by reasoning over a host-declared tool schema, populating structured tool arguments, chaining tool outputs across multi-turn trajectories, and signaling completion via a dedicated submit/terminate protocol. Rather than executing real-world actions itself, MagenticBrain delegates all execution to controlled host systems. It is co-designed with and optimized for MagenticLite, Microsoft's agentic application and harness, which is the recommended deployment vehicle and the configuration in which the model has been most thoroughly evaluated. Primary use cases include workflow planning and tool orchestration, file-system and terminal task coordination, delegation to specialized sub-agents (for example, Fara1.5 for web browsing or CUA for computer use), and enterprise automation scenarios that require deterministic tool boundaries, auditability, and human-in-the-loop oversight. MagenticBrain supports a 32,000-token context window covering the system prompt, tool schemas, user goal, and prior trajectory state, with text-only input and structured tool-call output. The primary public configuration is fine-tuned from the Qwen 3 14B base model. Post-training uses supervised fine-tuning on a curated recipe of agentic data spanning tool-calling and orchestration corpora (APIGen-MT, ToolACE, xLAM), file-system and containerized terminal tasks with executable tests, explicit sub-agent delegation traces, and reasoning and planning data, with shared JSON/XML normalization, context-length filtering, and think-block controls applied to reduce verbosity and looping. The training mix is aligned with how MagenticLite invokes the model in production, biasing the fine-tune toward stable, predictable behavior in that harness. Reinforcement learning is being actively explored to improve terminal-task robustness and delegation quality.