MiMo-V2.6

Xiaomi’s MiMo project is back with its seventh launch, and this one is bigger than a simple model drop. MiMo-V2.6 is the company’s open omnimodal model family, built for long-horizon agent work rather than quick one-off prompts.

The headline feature is breadth. MiMo-V2.6 Pro and Flash can handle text, images, audio and video, and both come with a 1M context window. That puts the emphasis on sustained tasks, where a model has to keep track of a long thread instead of forgetting the plot halfway through.

But Xiaomi didn’t stop at the models themselves. It is also publishing the technical report, RL environments and the training code behind the public post-training run. That matters because it gives outsiders something more useful than a polished demo: a chance to inspect how the system was shaped after pretraining.

This is the kind of release that says open models are moving past simple availability and into reproducibility. The model family is the product, sure, but the surrounding materials are what make MiMo-V2.6 feel like a public research effort rather than just another branded launch page.