MiMo V2.6 Pro

MiMo V2.6 Pro was released September 21, 2026 by Xiaomi's MiMo team, led by ex-DeepSeek researcher Luo Fuli: a 1.02-trillion-parameter sparse MoE with only 42B active per token, natively omnimodal across text, image, video, and audio, with a 1M-token context window — all under an MIT license. Xiaomi live-streamed the reinforcement-learning post-training on a public dashboard, a reported $2.62M across 750,000 trajectories in under six days, an approach the team calls 'You Only RL Once.'

It debuted atop the Artificial Analysis Intelligence Index for open-weight models at 46, tying Grok 4.7, with near-parity to Claude Opus 5 and GPT-5.6 Sol on DeepSWE v1.1 and OSWorld-Verified per Xiaomi's model card. Hosted pricing is $0.435/$0.87 per million tokens — roughly one-twentieth to one-sixtieth the cost per intelligence-index task of leading closed models. Caveats: most benchmark figures are Xiaomi's own, ExploitBench is a weak spot (47.9 vs Sol's 78.5), and some organizations will weigh the geopolitical dimension of a Chinese-lab model.