On September 22, 2026, Xiaomi's MiMo team officially released and open-sourced the MiMo-V2.6 series of large models through official channels: the flagship MiMo-V2.6-Pro and the efficient reasoning model MiMo-V2.6-Flash. Both are natively omnimodal, accepting text, image, video and audio input, under the MIT license — the weights can be downloaded, deployed and used commercially.
What a score of 46 means: No. 1 among open models
The headline figure comes from the third-party platform Artificial Analysis: MiMo-V2.6-Pro scored 46.32 on the Intelligence Index, surpassing Kimi K3 and Qwen3.8 Max to become the highest-scoring open-weight model at the time of release. Xiaomi's wording is carefully hedged: Pro performs on par with Claude Opus 5 and GPT-6 Sol across "most" agent benchmarks, not all of them. The leading closed models still hold a one-to-four-point edge, but it is the first time a Chinese open-source effort has touched the threshold of the world's top tier.
The training run streamed live for the world: six days, $3.47 million
The real story of this release is not the parameter count but the training. The MiMo-V2.6 series keeps the same pre-training architecture as the previous generation; the capability leap comes almost entirely from reinforcement-learning post-training. Xiaomi streamed the entire RL run on a public dashboard: cost, token throughput and per-step benchmark scores, all visible. Pro and Flash each completed 30 RL steps and roughly 750,000 trajectories in under six days, at a total cost of about $3.47 million. Outside observers even recorded a GPU-memory-triggered restart mid-run.
Not just the weights: the RL toolchain came along too
What deserves the most attention is the accompanying open release: the distilled small model MiMo-V2.6-Distill-Qwen-9B, more than 7,000 high-quality task environments, an end-to-end RL training framework and a lightweight, composable agent framework. Training environments, training code and verification tooling are all open.
Why it matters: three reasons to take this seriously
First, full multimodality becomes the standard configuration of an open Pro-class model for the first time; multimodal understanding is no longer a privileged gateway of closed models. Second, the "open hand" approach to scaled RL: publishing the training process, the cost and even the failures is both engineering confidence and a move to make reproducibility the new industry bar. Third, price: API pricing stays at V2.5 levels, with Pro's measured per-task cost at $0.13, just one-twentieth to one-sixtieth of some overseas models. Of course, 46 points does not mean sweeping superiority — Xiaomi's own tables show Pro trailing in some coding and security tests. The measured reading: the open camp has, for the first time, a Chinese model that can stand on the same stage as the top closed models, and its training methods and toolchain are open to the world.
