Xiaomi open-sources MiMo-V2.6 Pro and Flash for agents

Xiaomi’s MiMo team released the MiMo-V2.6 series on September 22, 2026, and posted the weights on Hugging Face. The drop is two natively omnimodal models—MiMo-V2.6-Pro as the flagship and MiMo-V2.6-Flash as the efficiency path—plus a technical report. The intended workload is long-horizon coding and general agents: long repositories, tool traces, and multi-session runs, not just chat.

Flash is the checkpoint with a published architecture card. It is a sparse mixture-of-experts model with 309 billion total parameters and 15 billion activated, a 1 million-token context window, and native text, image, video, and audio. Xiaomi’s Hugging Face collection groups MiMo-V2.6-Pro-RL, MiMo-V2.6-Flash-RL, and MiMo-V2.6-Distill-Qwen-9B together.

Vendor-reported agent numbers (not independently verified here) put Pro at 71.9 on DeepSWE v1.1, 53.1 on AutomationBench v1.0.6, and 89.9 on Terminal Bench 2.1. Flash is listed at 67.9, 52.3, and 87.6 on the same three. Xiaomi also cites a 46.32 Artificial Analysis Intelligence Index for Pro and says API pricing is unchanged from the V2.5 series.

The training story is one mixed reinforcement-learning run across coding, general agents, visual tasks, and cybersecurity, rather than separate per-domain jobs. Xiaomi says each of Flash and Pro completed 30 RL steps over roughly 750,000 trajectories in under six days, at about $0.85 million and $2.62 million respectively, and that it is open-sourcing the technical report, training environments, and RL code.

The models are available in AI Studio, MiMo Code, MiMo Desktop, the MiMo API Platform, and OpenRouter. Desktop ships its first official release with Pro and Flash built in, and adds MiMo-V2.6-Pro-UltraSpeed, described as up to 20× faster output at the same quality. Listed USD rates per million tokens are Flash $0.0028 / $0.14 / $0.28 (cache hit / miss / output), Pro $0.0036 / $0.435 / $0.87, and UltraSpeed $0.036 / $4.35 / $8.7.

Download MiMo-V2.6-Flash-RL or Pro-RL from Xiaomi’s Hugging Face collection, then read the launch page and the technical report on the Flash model card before you drop either checkpoint into a long agent trace.

Sources