MiMo V2.5 Base
V2.5 Base is the multimodal checkpoint before the final agent-training stages. It retains the vision and audio paths derived from the V2.5 architecture.
Research release
Multimodal pretraining precedes agent training
The shared training description separates multimodal pretraining from supervised agent tuning and reinforcement learning. This makes the base release useful for studying an alternative post-training recipe.MiMo-V2.5-Base official model card
Set a separate baseline for image and audio tasks
Practical guidance. Before adapting the checkpoint, record its text, image and audio behavior. A gain on one input type can conceal a loss on another. Keep a fixed set of mixed-modality tasks alongside the target-domain examples.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-V2.5-Base |
| Context window | 262,144 tokens Base-specific download table lists 256K. MiMo-V2.5-Base official model card |
| Size | 310 billion total parameters; 15 billion active per token.MiMo-V2.5-Base official model card |
| Licence | MIT.MiMo-V2.5-Base official model card |
| Context distinction | The download table gives Base 256K and the post-trained model 1M.MiMo-V2.5-Base official model card |
Related models
A multimodal model combining the V2 Flash language backbone with vision and audio encoders.
Launch date not confirmedDownloadable