Xiaomi MiMo's public reinforcement learning training runs stopped, completing a combined total of 60 steps

Article

Xiaomi MiMo goes public RL Stopped: two training runs burned through 3470000 USD Beating AI News Flash: the two reinforcement learning training runs publicly livestreamed by Xiaomi MiMo have now both stopped. Pro and Flash each last fully completed 30 training steps, with the two runs spending a total of 3474700 USD. Pro spent 2620700 USD, while Flash spent 854000 USD.

The gains bought with that spending are clear. On DeepSWE v1.1, Pro rose from 1 step 58.41 to 72.57, an increase of 14.16 points; Flash rose from 48.67 to 65.68, an increase of 17.01 points. However, both curves showed clear declines in between; scores did not rise continuously with longer training.

The other two dashboard evaluations also improved. Pro’s internal code evaluation rose from 57.54 to 65.43，AutomationBench, and from 45.2 to 53.1.

The training process was not smooth. Pro previously triggered GPU OOM due to imbalanced expert loads, and the training cluster and evaluator also lost connection and restarted; Flash was rerun from step 15 due to an infrastructure error. Later, Xiaomi also filtered out tasks that had become too easy for Pro and removed the cyber dataset exhibiting anomalous rollout patterns from Pro’s subsequent training.

Original link https://m.theblockbeats.info/flash/368235

Key points

Pro and Flash each last fully completed 30 training steps.
Total cost of the two runs 3474700 USD, including Pro cost 2620700 USD, Flash cost 854000 USD。
On DeepSWE v1.1, Pro increased from 58.41 to 72.57, while Flash increased from 48.67 to 65.68。
Pro's internal code evaluation increased from 57.54 to 65.43，AutomationBench from 45.2 to 53.1。
Pro once triggered GPU OOM due to imbalanced expert loads GPU OOM, and the training cluster and grader were also disconnected and restarted.
Flash was rerun from step 15 due to an infrastructure error.
Xiaomi filtered out tasks that had become too easy for Pro and removed cyber datasets exhibiting anomalous rollout rollout patterns.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Both training runs have stopped; evaluation scores increased overall, but declined at points in between, and training also encountered GPU memory, disconnection, and infrastructure issues.

Why it matters to readers

The results show that higher training costs were accompanied by improved evaluation performance, but the training process did not increase continuously and experienced infrastructure failures.

Reinforcement learning training can be understood as a training process that progresses by step; here, steps, costs, and evaluation scores are used to observe training progress.

Risks and unknowns

The information is sourced from breaking-news briefs and original links; no independent verification information was provided.
The specific reason for the training stop was not clearly stated.

Xiaomi MiMo's public reinforcement learning training runs stopped, completing a combined total of 60 steps

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center