Xiaomi confirms MiMo-V2.6 suffers from a looping issue: Reinforcement Learning (RL) worsened the bad behavior during training, with repairs totaling $90,000.

Beating AI Express, Insight: Xiaomi’s MiMo team has reviewed the repeated tool call issue following the launch of MiMo-V2.6. The model sometimes repeatedly invokes identical or highly similar tools, draining context without advancing task progress. In OpenCode, the share of responses with repeated tool calls for Flash and Pro once hit 1.02% and 0.54% respectively. The root cause lies in reinforcement learning (RL) reward design: training prioritizes rewarding "final task completion accuracy" but fails to adequately penalize inefficient in-process behaviors. The original rule only imposed penalties when single-round tool calls exceeded 32 times; repeated calls under 32 times incurred no penalty at all. As RL scale expanded, this "bad habit" was further reinforced. After reviewing training checkpoints, Xiaomi found the proportion of abnormal samples with single-round calls over 10 times in Flash rose from 11.1% at step 0 to 24.6% at step 20. The most direct fix would have been cutting the penalty threshold from 32 to 8 calls, then re-running ~20 MixRL steps, estimated at $2.31 million. Instead, Xiaomi trained a dedicated RL teacher to correct repeated calls, running 12 steps and ~7,000 samples, then integrating this capability back into Pro and Flash via MOPD. The full fix cost ~$90,000, just 4% of the full retraining plan, with other key benchmarks largely unchanged. The fixed MiMo-V2.6-Pro-MOPD and Flash-MOPD weights are now available, and the API has been switched to the new version with unchanged call names. Xiaomi will also reset the remaining quota for MiMo Desktop users’ current cycle.