MiMo V2.5
V2.5 adds vision and audio encoders to the V2-Flash backbone. It accepts text, images, video and audio and was trained for multimodal understanding and agent work.
The audio path builds on MiMo-Audio
The audio encoder starts from MiMo-Audio tokenizer weights. The card describes projector warmup, multimodal pretraining, supervised tuning, agent reinforcement learning and teacher distillation. The final context was extended to one million tokens.MiMo-V2.5 official model card
The separate MiMo-V2.5-DFlash repository offers a deployment package under MIT. Its model card does not provide a detailed operating guide. The package is listed with this release rather than counted as a new foundation-model generation.MiMo V2.5 DFlash package
Compare each input modality separately
Practical guidance. A model may be strong on text while missing important details in video or audio. Use paired cases with the same underlying facts in different input formats. Check the multimodal processor files along with the weights when reproducing a result.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiMo-V2.5mimo-v2.5MiMo-V2.5-DFlash |
| Context window | 1,000,000 tokens Post-trained V2.5 capacity. MiMo-V2.5 official model card |
| Size | 310 billion total parameters and 15 billion active parameters.MiMo-V2.5 official model card |
| Encoders | A 729-million-parameter vision transformer and a 261-million-parameter audio transformer.MiMo-V2.5 official model card |
| Licence | MIT.MiMo-V2.5 official model card |
| Original API | Retired on 21 October 2026 at 10:00 UTC+8; V2.6 Flash is the service replacement.MiMo model retirement and replacement schedule |
Related models
- MiMo V2.6 Flash
Xiaomi · MiMo V2.6
A multimodal agent model that accepts text, images, video and audio and returns text.
Launch date not confirmedAPI · Downloadable · Provider product - MiMo Audio 7B Instruct
Xiaomi · MiMo Audio
An instruction-tuned audio model for spoken dialogue, audio understanding and text-to-speech.
Launch date not confirmedDownloadable - MiMo V2.5 Base
Xiaomi · MiMo V2.5
V2.5 Base is the multimodal checkpoint before the final agent-training stages.
Launch date not confirmedResearch releaseDownloadable