Xiaomi Mimo 2.6 live post-training dashboard

Xiaomi's MiMo dashboard publishes logs from the reinforcement-learning runs (training in which a model is rewarded for completing tasks) behind its v2.6 Pro and Flash models. Its headline number, called avg@n, takes each prompt sampled at a training step, measures the share of that prompt's n attempts that succeed, and averages those shares across prompts. A second version leaves out attempts that failed for infrastructure reasons, which separates a model failing a task from the training machinery failing to finish an attempt. The dashboard defines these measures alongside reward, step timing and workload counts.

These figures describe the training process. The headline score concerns prompts sampled at each step, so it does not establish performance on a fixed independent test or a ranking against other models. The published status data now marks both runs as ended. The archive remains useful for seeing what the team monitored, including attempts lost to infrastructure failures and, on average, how many model updates occurred between generating an answer and using it for training.