Reinforcement learning runs boosted the performance of the MiMo-V2.6-Pro model on the DeepSWE v1.1 coding test. The model's score rose from 58.41 to 72.57, which is 1.43 points below the leading mark of 74 currently held by Astra, Gemini 3.8 Flash, and Opus 5.

The 72.57 result serves as a checkpoint from a live training process rather than a final score. This increase puts the MiMo-V2.6-Pro model in a similar performance bracket as the current top-tier AI assistants, reducing the gap between the pro model and the market leaders.