The models were trained via scaled RL using a Hybrid SWA architecture, with training costs of $2.6 million for Pro and $0.9 million for Flash. In a materials science case study, the Pro model shortened a research cycle from one month to 2 to 3 days, a 10x productivity gain. While the full framework was released, one reviewer noted that only 989 environments were used for the 9B distilled version. Xiaomi has already announced a new architecture called HySparse2 for the upcoming MiMo-V3 to improve memory efficiency and prefill speeds for agentic workloads.
Tuesday, October 20, 2026
Story
Xiaomi MiMo-V2.6 Takes No. 1 Open Weight Rank on Vals Index
AI25 days ago
A new suite of omnimodal intelligence models from Xiaomi now leads competitors in agentic tasks, with the Pro version reaching 72.57 on the DeepSWE benchmark. The MiMo-V2.6-Flash model holds the top position among all open-weight models on the Vals Index with a 59.6% score, narrowly beating the Pro version's 59.5%. The Pro edition is an MoE model featuring 1.02 trillion total parameters and 42 billion active parameters, and it scored 46 on the Artificial Analysis Intelligence Index. Xiaomi released the weights under an MIT license and provided the end-to-end training code and over 7,000 reinforcement learning (RL) environments.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.