Xiaomi open-sourced its MiMo-V2.6 model series, led by two natively omnimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, plus a speed-tuned inference mode called MiMo-V2.6-Pro-UltraSpeed. Along with the model weights, Xiaomi published the full technical report and the reinforcement-learning training code and environments behind the release.
The approach behind MiMo-V2.6 matters as much as the release itself. Xiaomi describes it as a recursive self-improvement path: instead of just scaling up parameter count, the company pours reinforcement-learning compute into hard, verifiable tasks so a model can keep sharpening its own abilities through repeated trial, feedback, and correction.
The proof, in Xiaomi's telling, is in the benchmark score. According to Xiaomi, MiMo-V2.6-Pro scored 46.32 on Artificial Analysis's Intelligence Index, ahead of Kimi K3 and Qwen3.8 Max, a result the company calls the strongest showing yet for an open-source model. Xiaomi held API pricing level with the outgoing V2.5 series, so the accuracy gain arrives without a price increase.
The bigger commitment is what Xiaomi handed over for free. Xiaomi says MiMo-V2.6-Flash and MiMo-V2.6-Pro each completed 30 reinforcement-learning steps over roughly 750,000 trajectories in under six days, at a cost the company puts at about $0.85 million and $2.62 million respectively. On DeepSWE, a long-horizon coding benchmark, scores rose from 48.8 to 65.68 for Flash and from 58.4 to 72.57 for Pro, evidence Xiaomi points to as proof the gains generalize past the exact tasks used in training.
Coding is the entry point, not the ceiling. Xiaomi frames MiMo-V2.6 as extending natural-language programming from software into what it calls Vibe World: turning a prompt into a playable 3D scene, controlling a robotic arm through visual feedback, drafting a slide deck or frontend interface, and scoring an orchestral piece it then converts into a working MIDI file on its own. In case studies Xiaomi published with the release, the model reportedly helped design a candidate material for capturing so-called forever chemicals and helped formalize a classic dynamical-systems proof.
You do not need to wait for a research preview to try any of this. MiMo-V2.6-Pro and MiMo-V2.6-Flash are already live in AI Studio, MiMo Code, and MiMo Desktop, and reachable through the MiMo API platform and OpenRouter. MiMo Desktop leaves early access with this release, and Xiaomi is rolling out the UltraSpeed mode for work that needs a faster response than the standard Pro model provides.
What's still unclear is how MiMo-V2.6 holds up once it leaves Xiaomi's own benchmarks and case studies, and whether independent developers pulling the open weights from Hugging Face and OpenRouter find it as capable in practice as the company's own numbers suggest.
