MiMo V2 Flash
V2 Flash is the text-model release that introduced the architecture later used by V2.5. It combines sliding-window attention, global attention and multi-token prediction.
A short local window is mixed with global layers
The card gives a 5:1 sliding-to-global attention ratio and a 128-token sliding window. Learnable attention-sink bias is part of the long-context design.MiMo-V2-Flash official model card
Treat attention design as an engineering tradeoff
Practical guidance. Compare retrieval accuracy and memory use across several input lengths. A lower cache requirement is useful only if the model still uses the evidence needed for the task. Keep local context settings in the comparison record.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | MiMo-V2-Flashmimo-v2-flash |
| Context window | 262,144 tokens Published V2 Flash support up to 256K. MiMo-V2-Flash official model card |
| Size | 309 billion total parameters and 15 billion active parameters.MiMo-V2-Flash official model card |
| Training | 27 trillion tokens, with a native 32K training sequence length before longer-context support.MiMo-V2-Flash official model card |
| Licence | MIT.MiMo-V2-Flash official model card |
| Original API ID | Stopped working on 30 June 2026 at 00:00 UTC+8.MiMo model retirement and replacement schedule |
Related models
A multimodal model combining the V2 Flash language backbone with vision and audio encoders.
Launch date not confirmedDownloadable- MiMo V2 Flash Base
Xiaomi · MiMo V2
V2 Flash Base is the pretrained 309B mixture-of-experts checkpoint.
Launch date not confirmedResearch releaseDownloadable