Xiaomi's MiMo-V2.6-Pro surpasses competitors in code arena

Xiaomi's MiMo-V2.6-Pro has made a strong comeback in the open-source large model rankings, achieving a score of 1628 in the WebDev test field. This places it among the top three models using the MIT license. The model outperforms Hy4-preview with a score of 1624 and surpasses its predecessor, MiMo-V2.5-Pro, which scored 1475. The performance marks a 153-point improvement in a single iteration. In the broader open-source weight model comparison, MiMo-V2.6-Pro trails Qwen3.8Flash Next by 7 points and is 46 points behind Kimi K3Max. The results are based on automated voting using a reward model trained on human preference data. As more human voting data accumulates, the final rankings will be refined. Xiaomi's achievement highlights its advancements in reinforcement learning and all-modal large model development. The model's strong debut underscores Xiaomi's ongoing commitment to AI innovation. The performance in the Code Arena demonstrates significant progress in code generation and web development tasks. The company's focus on technical breakthroughs continues to position it as a key player in the open-source AI landscape. The results reflect the growing competitiveness of Chinese-developed models in global benchmarks.