MiroThinker-v1.0-72B: Interactive Scaling as the Third Dimension Hits 81.9% on GAIA

HuggingFace / MiroMind AI·high signal

MiroThinker-v1.0-72B (open weights on HuggingFace) achieves 81.9% on the GAIA-Text benchmark via 'interactive scaling'—training the model to systematically leverage environment feedback and external information acquisition as a third scaling dimension beyond model size and context length. It supports 256K context and up to 600 tool calls per task, outperforming MiniMax-M2 (75.7%) by 6.2 points. Unlike test-time compute scaling that risks reasoning chain degradation in isolation, interactive scaling uses external corrections to keep trajectories on track.