Xiaomi released MiMo-V2.6 today, and the release notes read less like a routine update and more like a case study in how far reinforcement learning can stretch a training budget. Two models, a 1.02 trillion-parameter Pro and a leaner 309-billion-parameter Flash, both natively multimodal, both open-weight under MIT, both trained through a six-day RL run that Xiaomi streamed live.
 
The scores are the part that stands out. Pro hits 46 on the Artificial Analysis Intelligence Index, the highest mark any open-weight model has posted there, tying it with a proprietary rival that costs several times more to run per token. It also tops the field on a cybersecurity benchmark and adds computer-use skills the previous generation didn't have.
 
We go through the architecture, the training approach, the reward-hacking defenses Xiaomi built in, and where the model still falls short against closed-source leaders.
 
Read the full story: https://literouteracademia.com/2026/09/23/xiaomi-releases-mimo-v2-6-as-its-strongest-open-model-yet/