Xiaomi releases MiMo‑V2.6 Pro and Flash as open‑weights models
Two accounts put the larger model near the top of open-weight rankings, on different benchmarks, and one of those scores rests on automated votes.
Xiaomi has released MiMo-V2.6-Pro and MiMo-V2.6-Flash as open-source models, according to the tracking account testingcatalog, which said the larger model reached sixth place on the Artificial Analysis Intelligence Index. No post from Xiaomi appears in the material gathered for this story.
Arena, which runs the Code Arena leaderboard, said MiMo-V2.6-Pro debuted at about tenth overall on its WebDev board and about third among open-weights models carrying an MIT licence. Arena put the score at 1,628 points, a rise of 153 from MiMo-V2.5-Pro.
Arena said the number is an early AutoEval score, in which a reward model trained on its human preference data casts votes in place of live ones. Arena said it expects the score to move as human votes arrive. The two rankings measure different tasks.
The account eliebakouch said the team plans to publish about 7,000 items of reinforcement-learning training data along with the framework behind the model. That account also said the model and its technical report shipped less than a week after the final reinforcement-learning run started.
The account teortaxesTex put the run at five days on a cluster of roughly 10,000 GPUs and said the result is hard to dismiss as distillation from a larger model. That reading is one commentator's. Xiaomi has not confirmed the cluster size or the training timeline here.