Inside Xiaomi’s MiMo-V2.6 RL Dashboard

Xiaomi’s public MiMo-V2.6 dashboard made two reinforcement-learning post-training runs observable while they were underway. It showed more than a final score or model release, but its logs should be read as Xiaomi’s account of the Pro and Flash runs—not as a view of pretraining or every activity on its infrastructure.118

What did the dashboard show?

The Pro and Flash views tracked completed training steps, rollouts in progress, token counts, reward curves, task pass rates, step timings, accumulated costs and intermediate coding evaluations. Training-step and rollout-step counters could differ because rollout generation ran asynchronously with training updates.11822

The dashboard also exposed operational signals. Reports described selected restarts, GPU out-of-memory errors and dataset issues; its metric definitions distinguish ordinary task failures from attempts lost to infrastructure problems. Those disclosures are useful, but they do not establish that every incident was shown.1823

One early snapshot put both models at training step 10, with rollout step 16 underway. It reported about 2.2 billion tokens per Pro step and 2.6 billion per Flash step, alongside accumulated spending of roughly $741,000 and $322,000. Those were point-in-time figures, not final costs or stable daily rates.25

How were the training and grading runs configured?

Each update used 1,568 prompts with 16 attempts per prompt, or 25,088 potential trajectories. The asynchronous rollout setup let generation and task execution proceed alongside grading and model updates. The run mixed agent tasks—including coding, general-agent, visual and cybersecurity work—across multiple harnesses rather than training against one task type alone.41019

Grading went beyond a simple pass/fail test. Xiaomi’s described approach combined executable-test outcomes with task-specific rubrics and comparisons among attempts for the same task, allowing successful attempts to receive different quality assessments. Grading consumed compute too; an account of Xiaomi’s technical report puts it at about 12.7% of Pro’s RL cost.4447

The early description of “roughly 2 billion tokens per step” was a scale estimate, not a fixed quota. Later accounts report higher per-step token counts.41920

What do the rewards, DeepSWE scores and final costs mean?

Training reward describes the trajectories used for an update; it is not itself an independent capability test. The dashboard defines dynsam/avg@n as the average, across sampled prompts, of the fraction of attempts that succeed. Its _no_infra version excludes attempts that failed for infrastructure reasons. Comparing the two helps avoid mistaking a broken rollout for a model failure, though neither metric measures performance across all possible tasks.18

The early snapshot reported DeepSWE v1.1 scores of 62.24 for Pro and 60.77 for Flash. They were interim checkpoint evaluations. Xiaomi later reported final-run results of about 72.6 for Pro and 65.7 for Flash; those figures should still be understood in the context of Xiaomi’s evaluation setup, not treated as independently reproduced public-leaderboard results.251745

Xiaomi says both runs finished at 30 steps each in under six days, at approximately $2.62 million for Pro and $850,000 for Flash—about $3.47 million together for the reported RL runs. The stated total of roughly 750,000 trajectories is best read as per model: 25,088 potential trajectories per step multiplied by 30 steps equals 752,640 for each run. These reported expenditures do not include pretraining or all other model-development work.244445

What was distinctive—and what remains unknown?

Unlike a release that reveals only finished model artifacts, this dashboard exposed time-series training and operational data while the runs were in progress. Xiaomi has since announced Pro and Flash weights, a 9B distilled model, a technical report, more than 7,000 RL task environments, an end-to-end RL framework and composable mini-harnesses.115

That greater visibility has limits. The available sources do not independently verify uninterrupted, unfiltered live continuity or show all concurrent infrastructure activity. Nor does the existence of a released 9B distillation establish that undisclosed distillation traffic ran behind the displayed Pro and Flash jobs or was included in their reported costs.18152