On September 22, 2026, Xiaomi simultaneously announced the MiMo-V2.6 series through its official MiMo X account and WeChat public account, including two versions: MiMo-V2.6-Pro and MiMo-V2.6-Flash. The most noteworthy aspect of this update is not the parameter numbers, but how Xiaomi has transformed multimodal capabilities from "understanding images" into the foundation for Agent tasks—coding, computer use, 3D reasoning, and creation are all listed in one capability list.
Pro and Flash: A dual release with clear divisions of labor
Pro handles capability ceilings, while Flash handles high-frequency calls. The official emphasis is that both versions have undergone large-scale reinforcement learning training and are native multimodal. This "one strong, one fast" combination is familiar: currently, the common division of labor in agent infrastructure is planning large models and execution using small, fast models, which Xiaomi has directly integrated into its official product line. For developers, the selection issue shifts from "which model to use" to "how to combine":P RO handles planning and complex tasks requiring deep reasoning, while Flash handles high-frequency tool calls and execution steps. Both share the same multimodal understanding capabilities, with context and tool definitions reusable.
Full-modal plus large-scale reinforcement learning is Xiaomi's directional statement
MiMo-V2.6 is defined as a native full-modal model, meaning that beyond text and images, spatial understanding abilities like 3D reasoning are also integrated into the same training system. The official focus on "scaled reinforcement learning" is at the core, indicating that this generation's focus is not on competing on parameter scale, but on aligning multimodal capabilities with real tasks, especially in Agent scenarios that require multi-step operations and tool calls. For model researchers, MiMo-V2.6 is another example of the "scaling reinforcement learning" approach: as multimodal capability improves from pre-training to post-training, the focus of evaluation should shift from single-point benchmarks to true completion rates of long tasks. Model weights have been simultaneously released through official channels, and whether the promoted task performance is delivered can be tested and verified.
Who should follow up?
Teams working on agent applications can focus directly on computer use and coding capabilities and conduct a hands-on comparison with the Flash version; Developers interested in the open-source ecosystem can verify actual performance after weight opening; Teams still selecting multimodal foundations can treat the Pro/Flash division as a ready-made selection framework for reference. To understand the previous version history of the MiMo series, you can check out the Xiaomi MiMo special topic.
