Xiaomi open-sources MiMo-V2.6 models after intensive reinforcement learning runs
Xiaomi open-sourced its multimodal MiMo-V2.6-Pro and Flash models on September 22, spending $2.62 million and $850,000 on 30-step reinforcement learning runs across 750,000 trajectories each.
Why it mattersXiaomi claims Pro leads open-source models with a 46 on the Artificial Analysis index over Kimi K3, but we still need public blind Arena votes to verify it.
Very likely92%▲ from 1% a week ago
How we got here
- Xiaomi’s team studied how far reinforcement learning could scale and placed MiMo-V2.6 into a training run. Yahoo
- Coverage highlighted Pro’s Artificial Analysis score and detailed the series’ reinforcement-learning training. Winbuzzer
- Xiaomi announced that its invitation-only beta program would end in one week. Mi
What's likely next
- Xiaomi said its invitation-only beta program would end one week after its September 22 announcement—around September 29.
All coverage
The coverage
Something wrong or missing? Tell us.
More in Tech
CDC logs 3,659 measles cases as Pennsylvania reports four deaths
Will the U.S. record more than 5,000 measles cases in 2026? ↑4k: 98%
ElevenLabs launched Eleven v4 speech models
Which text-to-speech AI will rank highest this week? ElevenLabs leads: Very likely (88%) ▲ from 2%
Trump says AI will be renamed super intelligence in all US documents
Trump renaming AI by Sept 30: Unlikely (26%) ▲ from 8%
NWS reports falling Mississippi River stages upstream
SpaceX runs Flight 14 wet dress rehearsal ahead of first orbital attempt
Humans colonizing Mars by Jan 1, 2050: Toss-up (58%) ▲ from 20%