Xiaomi released the MiMo-V2.6 family under the MIT licence. Pro is a mixture of experts with 1.02 trillion parameters, 42 billion of them active per token. Flash is 309 billion total and 15 billion active. There is also a 9 billion parameter distillation of Qwen. All take text, images, video and audio, with a 1M token context. The unusual part is what shipped alongside the weights: more than 7,000 reinforcement learning environments and the framework that runs them. Most open releases give you the model. This one gives you the machine that trained it.