microsoft/MagenticBrain
MagenticBrain is a 14B-parameter orchestration model from Microsoft Research AI Frontiers, supervised fine-tuned from Qwen3-14B and further trained with reinforcement learning. It specializes in planning multi-step tasks, calling declared tools via structured JSON, and coordinating sub-agents, rather than acting as a general-purpose chat model. With a 32K context length, it is optimized for agentic task coordination within a host system like MagenticLite, handling file operations, terminal tasks, and sub-agent delegation.
Loading preview...
MagenticBrain: An Orchestration-First Agent Model
MagenticBrain, developed by Microsoft Research AI Frontiers, is a 14-billion parameter model specifically designed for agentic orchestration. Built upon the Qwen3-14B architecture, it has undergone extensive supervised fine-tuning and reinforcement learning on diverse agentic data, including function-calling corpora, file-system trajectories, and sub-agent delegation traces.
Key Capabilities
- Orchestration-first post-training: Specialized for planning, tool selection, multi-turn tool chaining, and delegating tasks to sub-agents, distinguishing it from general-purpose chat models.
- Structured JSON tool calls: Processes tool schemas passed at inference time and selects only from declared tools, never inventing new ones.
- Sub-agent delegation: Trained with explicit handoff traces for delegating computer-use tasks to sub-agents like Fara1.5-9B.
- Submit-to-terminate protocol: Utilizes a dedicated
submitsignal as a protocol token to indicate task completion, enabling clear end-of-task indicators for the host harness. - 32K context length: Provides ample capacity for system prompts, tool schemas, and multi-turn trajectory states.
Intended Use Cases
- Tool orchestration and workflow planning: Ideal for decomposing goals into steps, selecting tools, populating structured arguments, and chaining outputs across multiple turns.
- Agentic task coordination: Excels at managing file operations, terminal and code execution, and system actions through host-owned tools, including delegation to sub-agents.
- Research and experimentation: Suitable for benchmarking agentic reasoning, studying delegation patterns, and serving as an orchestrator in multi-agent stacks.