MiMo 7B RL Zero
RL Zero starts reinforcement learning directly from MiMo 7B Base. It is a distinct training experiment from MiMo 7B RL, which starts from a supervised-tuned model.
Research release
The release exposes two routes from the same base
Xiaomi publishes RL Zero alongside the base, SFT and final RL models. Its study uses these stages to examine how much reasoning capability is already present after pretraining.MiMo-7B-RL-Zero official model card
Measure format and correctness separately
Practical guidance. A direct-RL model may solve a task while presenting the result in an inconvenient form. Compare answer correctness, instruction following and stopping behavior separately. This is a research checkpoint, so the training comparison is its main value.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-7B-RL-Zero |
| Training path | Base → reinforcement learning, without the separate SFT checkpoint in between.MiMo-7B-RL-Zero official model card |
| Licence | MIT.MiMo-7B-RL-Zero official model card |
Related models
- MiMo 7B Base
Xiaomi · MiMo 7B
MiMo 7B Base was trained from scratch, with reasoning treated as a pretraining objective rather than added only after pretraining.
Launch date not confirmedResearch releaseDownloadable - MiMo 7B RL
Xiaomi · MiMo 7B
A 7B reasoning model trained with reinforcement learning for mathematics and code problems.
Launch date not confirmedDownloadable