MiMo V2.6 Flash
MiMo V2.6 Flash is Xiaomi's smaller current multimodal agent model. Its hosted API accepts text, image, video and audio input and returns text.
The update targets a behavior that final scores can miss
Xiaomi found that a model could issue repeated tool calls while appearing to make progress. Its published correction used a specialized teacher, then merged that behavior through MOPD. The downloadable MOPD release is distinct from the original RL checkpoint.MiMo V2.6 tool-call repetition investigation
Compare end-to-end work with Pro
Practical guidance. Flash is a useful comparison point for frequent requests and repeated agent work. Use the same task environment for both models. Count retries, failed actions and the time to an accepted result, rather than comparing token speed alone.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | API, Downloadable, Provider product |
| Model identifiers | mimo-v2.6-flashMiMo-V2.6-Flash-MOPD |
| Context window | 1,000,000 tokens Hosted Flash context limit. MiMo current model catalogue |
| Maximum output | 131,072 tokens Hosted maximum output, stated as 128K on the product page. MiMo current model catalogue |
| Published weights | The MOPD checkpoint has 309 billion total and 15 billion active parameters, under MIT.MiMo-V2.6-Flash-MOPD official model card |
| Architecture | The card lists a 681-million-parameter vision encoder and a five-layer speculative decoder.MiMo-V2.6-Flash-MOPD official model card |
| Service update | The current API includes the September 2026 tool-repetition correction without a change to its model ID.MiMo V2.6 tool-call repetition investigation |
Related models
- MiMo V2.6 Pro
Xiaomi · MiMo V2.6
MiMo V2.6 Pro accepts text, images, video and audio, then returns text.
Launch date not confirmedAPI · Downloadable · Provider product A multimodal model combining the V2 Flash language backbone with vision and audio encoders.
Launch date not confirmedDownloadable- MiMo V2.6 Flash RL
Xiaomi · MiMo V2.6
A smaller MiMo V2.6 mixture-of-experts model trained for multimodal understanding and long-context agent work.
Launch date not confirmedResearch releaseDownloadable