MiMo-V2.6-Pro and V2.6-Flash
Omni-modal 1.02T (Pro) and 310B (Flash) MIT-licensed models; Artificial Analysis rated Pro 46, top of open-weight models at launch.
One mixed RL run across coding, agents, vision and cybersecurity, with groupwise agentic grading and async GRPO. Xiaomi puts RL cost at about $2.62M (Pro) and $850K (Flash). 1M context; Pro $0.435/$0.87 per million tokens; a 10x-price UltraSpeed edition runs about 20x faster.
- Date
- Monday, 21 September 2026
- Lab
- Xiaomi (MiMo)
- Kind
- open-weights
- Access
- open weights
- Price
- $0.435 input / $0.87 output per M tokens (Pro), 2026-09
Figures
| Measure | Value | Measured by |
|---|---|---|
| Artificial Analysis Intelligence Index (Pro) | 46 top open-weight at launch; GLM-5.3 45, Kimi K3 44 | independent |
| RL training cost (Pro / Flash) | $2.62M / $850K reported by TestingCatalog and Winbuzzer | company |
| DeepSWE v1.1 (Pro) | 58.4 to 72.57 V2.5-Pro to V2.6-Pro | company |
HF weights 2026-09-21 UTC and OpenRouter listings the same day. AA index and cost figures come from press write-ups, not Xiaomi's post, which I could not open.
Sources
- huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL
- www.testingcatalog.com/xiaomi-open-sources-mimo-v2-6-pro-and-flash-models/
- winbuzzer.com/2026/09/24/xiaomi-mimo-v2-6-downloadable-ai-low-cost-apis-a004-xcxwbn/
This record was checked against its sources on 6 October 2026. How we check