Mimo V2.6 Pro: The Intelligence Leap We Didn't Expect

Xiaomi just released MiMo V2.6, and V2.6 Pro is now the number one open weights model on Artificial Analysis at 46, up from 26 for V2.5 Pro. And yet the architecture, the 1.02 trillion parameters, the 42 billion active, and even the price per token are exactly the same as before. So what changed is a single reinforcement learning run that Xiaomi streamed live, costs and crashes included, for about $3.47 million, eleven days after Anthropic accused them of distilling Claude. As we look at what that receipt actually covers and what it doesn't, the two Terminal Benches that make it look both frontier and benchmaxxed, the hallucination rate nobody is talking about, and what it takes to run it yourself, the real question is this. If a few dozen people and a few million dollars of RL can close this much of the gap, where exactly is the moat now? Timestamps 00:00 MiMo V2.6 Pro Changes Everything 01:16 The Model Barely Changed — So What Did? 05:39 The $3.47M Reinforcement Learning Run 10:41 The Anthropic Distillation Accusation 13:21 Benchmarks: Frontier Model or Benchmaxxed? 16:41 Can You Actually Run MiMo V2.6? 17:28 Why Xiaomi Is Building AI #xiaomi #mimo #openweights #llm #ai