MiMo V2 Flash Base
V2 Flash Base is the pretrained 309B mixture-of-experts checkpoint. It is published separately from the agent-tuned V2 Flash release.
Research release
The published comparison includes base-model tasks
The card reports base-model evaluations, including long-context retrieval. These results belong to the base checkpoint; final agent results should be read with the corresponding post-trained release.MiMo-V2-Flash-Base official model card
Use a base model when the training stage matters
Practical guidance. This checkpoint is useful for comparing adaptation methods from a shared starting point. For an end-user assistant, inspect the final tuned model first. Do not treat the base checkpoint as a ready-made replacement for it.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Access | Downloadable |
| Model identifiers | MiMo-V2-Flash-Base |
| Context window | 262,144 tokens Base download table lists 256K. MiMo-V2-Flash-Base official model card |
| Active parameters | 15 billion per token.MiMo-V2-Flash-Base official model card |
| Licence | MIT.MiMo-V2-Flash-Base official model card |
| Backbone | Hybrid sliding-window and global attention with a multi-token prediction module.MiMo-V2-Flash-Base official model card |
Related models
- MiMo V2 Flash
Xiaomi ยท MiMo V2
A text model combining sliding-window attention, global attention and multi-token prediction.
Launch date not confirmedDownloadable