Xiaomi's MiMo-V2.6 Models Leverage Claude for Performance Boost

Xiaomi's MiMo-V2.6-Pro model has achieved an Intelligence Index Score of 46 points, significantly boosted by training data sourced from Anthropic's Claude. The MiMo-V2.6-Pro, with 1.02 trillion parameters, incurred a training cost of $2.62 million, while its counterpart, the MiMo-V2.6-Flash, cost $0.85 million to train. Xiaomi reported that the DeepSWE score for MiMo-V2.6-Pro improved from 58.4 to 72.6, while MiMo-V2.6-Flash's score rose from 48.8 to 65.7, indicating substantial enhancements in coding capabilities. The models were trained using 7,000 tasks and 1,000 music composition tasks, reflecting a comprehensive training regimen.

Anthropic tracked 400,000 exchanges involving Xiaomi's models over a 20-day period in March and April 2026, contributing to the performance improvements through expanded reinforcement learning techniques. Xiaomi's approach included recording user conversations with its MiMo models and utilizing them to extract training data from Claude, as stated by Anthropic. This collaboration underscores the growing integration of advanced AI models in consumer applications, with Xiaomi's MiMo models now positioned as leaders in the open model landscape. The total exchanges generated by seven Chinese labs reached 190 million, indicating a robust ecosystem of AI development in the region.