Xiaomi announced and open-sourced its MiMo-V2.6 series on 2026-09-22: two natively omnimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, that accept text, image, video and audio. The Hugging Face model card for the Pro model lists 1.02 trillion total parameters with 42 billion activated per token, a 1 million token context window, and an MIT license; the Pro-RL and Flash-RL weights are published on Hugging Face and ModelScope. Alongside them Xiaomi released MiMo-V2.6-Distill-Qwen-9B, a technical report, and an UltraSpeed mode it says delivers up to 20 times faster inference on its own desktop client and API.
Xiaomi’s headline claim is that MiMo-V2.6-Pro scores 46 on the Artificial Analysis composite intelligence index, which it describes as the strongest open-source model available, ahead of Kimi K3 and Qwen3.8 Max. On DeepSWE v1.1 the company reports Pro improving from 58.4 to 72.6 and Flash from 48.8 to 65.7 relative to the previous generation, and it says Pro is comparable to Claude Opus 5 and GPT-5.6 Sol on Design Arena. API pricing is unchanged from the V2.5 series.
The release note is unusually specific about the reinforcement-learning run. Xiaomi says it used more than 7,000 RL task environments covering software engineering, vulnerability reproduction, knowledge-intensive work and web design, generated about 750,000 trajectories, and spent roughly 850,000 dollars of RL compute on Flash and 2.62 million dollars on Pro. For the 9B distilled model it reports SWE-bench Verified rising from 61.1 to 66.2 and Terminal Bench 2.1 from 37.1 to 52.8.
Why it matters: a trillion-parameter omnimodal model under a permissive MIT license, from a phone and car maker rather than a dedicated lab, extends the run of Chinese open-weight releases competing near the proprietary frontier, and the disclosed RL cost figures are a rare public data point on post-training spend. What it does not show: the AA index figure and the benchmark gains are reported by Xiaomi, the RL dollar figures cover the RL stage only and not pretraining, and open weights at 1.02 trillion parameters are open in licence but still require data-center hardware to run.