MiroThinker-v1.5-30B: Advanced Search Agent for Tool-Augmented Reasoning

MiroThinker v1.5-30B, developed by miromind-ai, is a leading search agent engineered to enhance tool-augmented reasoning and information-seeking. Unlike traditional models that primarily scale by size or context length, MiroThinker introduces interactive scaling, a novel approach that systematically trains the agent to manage deeper and more frequent interactions with its environment. This method leverages environment feedback and external information acquisition to refine trajectories and correct errors, leading to predictable performance improvements across various benchmarks.

Key Capabilities

• 256K Context Window: Supports long-horizon reasoning and deep multi-step analysis.
• Extensive Tool Use: Handles up to 400 tool calls per task, a substantial improvement over prior open-source research agents.
• Interactive Scaling: Improves performance through dynamic agent-environment interactions.
• Strong General-Research Performance: Achieves competitive results on benchmarks like HLE-Text (39.2%), BrowseComp (69.8%), BrowseComp-ZH (71.5%), and GAIA-Val-165 (80.8%), setting a new world-leading BrowseComp performance.

Good for

• Complex, multi-step research tasks requiring extensive information retrieval.
• Applications demanding deep analytical capabilities and long-horizon reasoning.
• Developing agents that require frequent interaction with external tools and environments.
• Scenarios where error correction and trajectory refinement through environmental feedback are crucial.