The Xiaomi MiMo team’s review found that MiMo-V2.6 had a problem with repeated tool calls that consumed context without advancing the task, and fixed it through targeted RL. The fixed weights are now available, and API has switched to the new version.
Xiaomi's MiMo team publicly stated that the next generation MiMo-V3 will adopt the HySparse2 architecture to reduce long-context Agent prefill computation and KV Cache usage. Tests in the article state that with a million-token input, prefill compute is reduced to approximately the original 1/5。
A developer said the MiMo Code installer contains an undisclosed module capable of traversing and packaging Git codebases, but no evidence has yet been found that the function was called or that a complete codebase was uploaded.
Xiaomi releases and open-sources the MiMo-V2.6 Pro and Flash native multimodal models, claiming that Pro achieved AA on the 46 Comprehensive Intelligence Index.