MiMo 7B RL
MiMo 7B RL applies reinforcement learning after the SFT stage. It targets mathematics and code problems with answers that can be checked by rules.
The rollout engine reduces idle work
Xiaomi's training design combines continuous rollouts, asynchronous reward calculation and early termination. The multi-token prediction layers are frozen during RL after earlier training.MiMo-7B-RL official model card
Check transfer outside verifiable puzzles
Practical guidance. Strong mathematics and coding results do not establish equal performance on open-ended writing or factual research. Compare the model on the intended task and inspect how it handles uncertainty when no automatic checker is available.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiMo-7B-RL |
| Training set | The release describes 130,000 mathematics and code problems for reinforcement learning.MiMo-7B-RL official model card |
| Reward design | Rule-based accuracy rewards, with data resampling for easier problems.MiMo-7B-RL official model card |
| Licence | MIT.MiMo-7B-RL official model card |
Related models
- MiMo 7B RL 0530
Xiaomi · MiMo 7B
An updated MiMo 7B reasoning model trained with an expanded supervised dataset and a longer reinforcement-learning window.
Launch date Downloadable - MiMo 7B RL Zero
Xiaomi · MiMo 7B
A MiMo 7B reasoning checkpoint trained with reinforcement learning directly from the base model.
Launch date not confirmedResearch releaseDownloadable - MiMo 7B SFT
Xiaomi · MiMo 7B
A supervised instruction-tuned MiMo 7B checkpoint used before reinforcement learning.
Launch date not confirmedResearch releaseDownloadable