Xiaomi's MiMo team released and open-sourced the MiMo-V2.6 series on September 22, launching the flagship MiMo-V2.6-Pro alongside the efficiency-focused MiMo-V2.6-Flash at the same time, with web, API and Hugging Face weight access all opened simultaneously. Both are natively omni-modal — text, images, video and audio all feed directly into the model — and the official focus areas are agentic tasks like coding, computer use, 3D scenes and audio-visual content creation.
The same day, Xiaomi also released a high-throughput version of Pro called Pro-UltraSpeed, saying it can push output speed up to 20 times faster at equivalent quality; this tier is being rolled out gradually. The desktop client, MiMo Desktop, has moved to a full release and now comes with a subscription.
Score and standing
Third-party evaluator Artificial Analysis gives MiMo-V2.6-Pro a composite Intelligence Index of 46, the highest among open-weight models, ahead of Kimi K3 and Qwen3.8 Max. According to leaderboards cited by media outlets, that score puts it in the same band as Claude Opus 5 and GPT-5.6 Sol, though it still lags behind Claude Fable 5.1 and GPT-6 Astra.
Pricing hasn't changed, carrying over the API rates from the V2.5 series. Artificial Analysis measured a per-task cost of $0.13; Xiaomi's own claim is that Pro costs only 1/20 to 1/60 of overseas models at an equivalent intelligence level. That multiplier comes from the vendor's own framing, and public materials don't fully specify which models it's being compared against or how the tasks were weighted.
Placing it next to April's generation
MiMo-V2.5-Pro was open-sourced at the end of April, specced as a 1.02-trillion-total-parameter, 42-billion-active-parameter MoE model with a 1-million-token context window. Back then Xiaomi's pitch was token efficiency, benchmarking ClawEval against DeepSeek-V4-Pro. By July, V2.5 had taken the top spot for both monthly and weekly call volume on OpenRouter, driven mainly by price.
V2.6-Pro's disclosed parameter scale is still 1.02 trillion total, 42 billion active — the same order of magnitude as the previous generation. The changes are concentrated in post-training. Lead researcher Luo Fuli called this round “one of the largest single reinforcement learning training runs any open-source model team has carried out so far,” saying the research and engineering difficulty exceeded what she encountered working on DeepSeek-R1.
According to training figures disclosed by media outlets, Pro and Flash each ran 30 steps, accumulating roughly 750,000 trajectories combined, at a total cost of about $3.5 million over a six-day period. Compared with pretraining runs that routinely cost hundreds of millions of dollars, that's not a large sum — but for a single post-training run, it's still unusual among open-source teams.
Strung together from April to September, Xiaomi's path looks fairly clear: lock in the base model's scale first, use price to win over developer call volume, then pour money and compute into reinforcement learning to push up the score. That score step has worked; what comes next is whether call volume can hold up without further price cuts.
Training method and what shipped alongside it
Luo Fuli described the team using a method called MixRL, which trains medium-difficulty, verifiable coding and agentic tasks jointly; games, 3D and subjective-evaluation tasks — which are hard to verify or run very long — are trained separately, with MOPD used at the end to merge the different capability tracks.
What's been open-sourced this time goes beyond the weights. Xiaomi has also released a Qwen model distilled from MiMo's training trajectories, roughly 7,000 diverse training environments, and the complete reinforcement learning framework. For teams looking to reproduce the work or run post-training on their own models, the environments and framework may end up being worth more than the weights themselves.
Luo Fuli also said on her personal account that MiMo-V2.6 is “already the top open-source model.” That claim currently rests on Artificial Analysis's composite index; rankings on individual leaderboards haven't all come out yet, and where it lands on code, image or voice sub-scores — and which leaderboard gets cited — could tell a different story. Xiaomi also hasn't said whether V2.6's open-source license carries over the same terms as the previous generation.
Sources: Xiaomi MiMo official release page, Artificial Analysis Intelligence Index leaderboard, CocoLoop, Sohu Tech. Intelligence Index scores and per-task cost are verified against Artificial Analysis's figures; training cost and step counts follow media disclosures.