MiMo-V2.6-Pro

Xiaomi's official MiMo release notes announce the open-source release of the MiMo-V2.6 series, including Pro and Flash variants as native fully multimodal models. The release describes six days of live RL training focused on recursive self-improvement via verifiable complex tasks, and reports that MiMo-V2.6-Pro scores The notes state the V2.6 series adopts the same API pricing as the V2.5 series, positioning MiMo-V2.6-Pro at roughly 1/20 to 1/60 the price of comparable overseas models at equivalent intelligence. MiMo-V2.6 is described as one of the domestic open-source models receiving the largest RL compute allocation to date, with

Latent Space's AINews coverage notes MiMo-V2.6-Pro is Xiaomi's most capable model to date and the new top open-weights entry on the Artificial Analysis Intelligence Index at 46, landing at $0.13 per task on the Pareto frontier. The article highlights Xiaomi's unusual transparency during final RL training runs, with the The piece frames MiMo-V2.6-Pro as crowning a new Chinese frontier lab, noting Xiaomi is not traditionally among the six Chinese AI Tigers. A separate UltraSpeed variant is mentioned as delivering up to 20x faster output at the same quality for latency-sensitive users. Total training cost is cited at approximately $3M,

LLM-Stats ranks MiMo-V2.6-Pro at 18 on its composite LLM Stats Score with a blended price of $0.46/M tokens and a score of 50.0, placing it between DeepSeek-V4.1-Flash (51.4) and Claude Opus 5.5 (59.7) on a log-scale cost-efficiency chart. Conversation-depth analysis reports stable scores around 13.1–13.4 across turns Benchmark entries name MiMo-V2.6-Pro explicitly with sourced scores from mimo.xiaomi.com, including a GDPval-AA Elo of 1670/3000, CyberGym at 0.94/1, and Terminal-Bench 2.1 referenced on the model page. The page presents these as the model's official scorecard alongside other tracked benchmarks.

Forkast reports Xiaomi released the MiMo-V2.6 series on September 22, 2026, with MiMo-V2.6-Pro scoring 46 on the Artificial Analysis Intelligence Index v4.3, tying Grok 4.7 and setting a new open-weight ceiling. The team, led by former DeepSeek engineer Luo Fuli, ships a frozen-router Mixture-of-Experts architecture wi The piece cites vendor-reported benchmarks for MiMo-V2.6-Pro including 71.9% on DeepSWE v1.1 (behind Claude Opus 5's 74.0%), 89.9% on Terminal Bench 2.1 (ahead of Claude Opus 5 and GPT-5.6 Sol), 94.0% on CyberGym (leading all reported models), and 53.1% on AutomationBench v1.0.6. The article notes these are vendor-repo

Flowtivity reports Xiaomi's MiMo-V2.6 series released September 22, 2026 in three variants: MiMo-V2.6-Pro (1.02T total / 42B active), MiMo-V2.6-Flash (309B total / 15B active), and MiMo-V2.6-Pro-UltraSpeed (a latency-optimized build). MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index, the highest ev According to Xiaomi's launch claims, MiMo-V2.6-Pro costs roughly 1/20 to 1/60 of leading international models at comparable intelligence. Both Pro and Flash ship under the MIT license, which the article highlights as enabling unrestricted commercial use by businesses for agentic workloads spanning tool use, terminal wo