FULL STORY

Xiaomi MiMo-V2.6: From RL Training to Open-Source Top

Xiaomi completed RL training for MiMo, boosting DeepSWE to 72.57, then open-sourced the MiMo-V2.6 series under MIT license, topping open-weight leaderboards with community benchmarks confirming major gains.

2026-09-21 ~ 2026-09-23 · 3 episodes · 43 posts

Episode 1 · Xiaomi MiMo Finishes RL Training, DeepSWE Score Jumps to 72.57 (2026-09-21, 2 posts)

Xiaomi's MiMo completed reinforcement learning training, boosting its DeepSWE score from 58.41 to 72.57—near the 74 record—while MiMo-V2.6 emerged as the latest open-source model to publicly document its RL training.

Episode 2 · Xiaomi Open-Sources MiMo-V2.6, Topping Open-Weights Leaderboard with Livestreamed RL Training (2026-09-22, 39 posts)

On September 22, Xiaomi open-sourced the MiMo-V2.6 model family on Hugging Face, comprising the flagship MiMo-V2.6-Pro and the lightweight MiMo-V2.6-Flash, both built on a MoE architecture with native omni-modal support and released under the MIT license. The Pro version topped the Artificial Analysis Intelligence Index for open-weight models with a score of 46, a big jump from its predecessor MiMo-V2.5-Pro's 26, making it the open-source community's most talked-about event of the day.

Confirmed

  • Two models released: MiMo-V2.6-Pro with 1.02T total parameters and 42B activated; MiMo-V2.6-Flash with 310B total parameters and 15B activated (cross-confirmed by multiple posts including @natolambert)
  • The Pro version scored 46 on the Artificial Analysis Intelligence Index, the top-ranked open-weight model; the predecessor MiMo-V2.5-Pro scored 26 (@ArtificialAnlys)
  • There is also an RL-trained MiMo-V2.6-Pro-RL variant, released alongside Flash with downloadable weights (@Bestlife73)
  • The series is open-sourced under the MIT license (@victormustar)

Why it matters

  • A phone maker's model topping a third-party intelligence index among open-weight models, and directly open-sourcing trillion-parameter-scale weights, has a real impact on the open-source landscape
  • The score jumped from 26 to 46 between generations—a striking leap—and @victormustar said it set a new record for open models on the Intelligence Index

19 more related posts →

Episode 3 · Xiaomi's MiMo-V2.6 Pro tested: big gains in coding, 3D games and multimodal (2026-09-22, 2 posts)

Hands-on tests of Xiaomi's open-source MiMo-V2.6 Pro cover coding, frontend, 3D game development and multimodal tasks, with Agentic Coding scores roughly tripling over the previous version and peak speeds hitting 900tps.