The Pro and Flash versions of the newly released MiMo-V2.6 multimodal AI model family outperform competing open-weight systems in intelligence and efficiency. MiMo-V2.6-Flash is the highest-ranked open-weight model on the Vals Index with a score of 59.6%, while the Pro version leads on the Artificial Analysis Intelligence Index with a score of 46. The Pro model uses a mixture-of-experts architecture with 1.02T total and 42B active parameters, costing $0.13 per intelligence index task.
Xiaomi utilized recursive self-improvement and scaled reinforcement learning across 30 training steps, spending $2.62M on the Pro model and about $850,000 on the Flash model. Along with the model weights under the MIT license, the company open sourced over 7,000 RL task environments and its training framework. The Flash model's cost is $0.20 per Vals Index test, roughly 1% of the price for Claude Opus 5.5. Xiaomi is already transitioning to a 'HySparse2' architecture for MiMo-V3 to improve memory efficiency.
Key sources
- SOURCE@valsai“MiMo V2.6 Flash (Xiaomi) - It scored 59.6% on the Vals Index and ranks #1 among open-weight models”x.com
- SUPPORT@valsai“MiMo V2.6 Pro (Xiaomi) - It scored 59.5% on the Vals Index and ranks #2 among open-weight models”x.com
- SUPPORT@artificialanlys“MiMo-V2.6-Pro is an MoE model with 1.02T total parameters and 42B active parameters”x.com
- SUPPORT@tim_dettmers“Around 250 tok/s decode 2.6 ktok/s prefill”x.com
- SUPPORT@theo“2.6 Pro in particular is doing well in some pretty hard tasks”x.com
- SUPPORT@maxforai“Flash 花了约 85 万美元，Pro 花了约 262 万美元”x.com
- SUPPORT@rasbt“average DeepSWE pass@1 accuracy on held-out harnesses improved from approximately 50% -> 66%”x.com
- SUPPORT@nrehiew_“pacing the frontier using only 30 steps”x.com