MiMo 7B Base
MiMo 7B Base was trained from scratch, with reasoning treated as a pretraining objective rather than added only after pretraining. It is the starting point for Xiaomi's original SFT and RL releases.
Research release
Multi-token prediction is part of the training
The card describes multi-token prediction as an additional objective. The release also publishes separate post-training paths from this base checkpoint.MiMo-7B-Base official model card
The MiMo-7B-MTPs repository supplies two additional pretrained prediction layers under MIT. Xiaomi says its released 7B models each use one prediction layer. The authors had not evaluated the post-trained models with both additional layers, so their presence is not evidence of a verified speed improvement.Additional pretrained MiMo 7B prediction layers
Use the base to isolate a training change
Practical guidance. Choose this checkpoint when comparing two adaptation methods from the same starting point. For a user-facing assistant, instruction following must be evaluated separately. A base model's continuation ability is not a chat-product contract.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-7B-Base |
| Pretraining data | About 25 trillion tokens, using a three-stage data mixture.MiMo-7B-Base official model card |
| Licence | MIT.MiMo-7B-Base official model card |
Related models
- 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