K2 Think V2

70B dense general-reasoning model from MBZUAI's Institute of Foundation Models, post-trained from K2-V2 Instruct via two-stage RLVR (Reinforcement Learning with Verifiable Rewards) on the improved Guru dataset. Released January 27, 2026 as the UAE's fully sovereign reasoning system — pre-training, post-training, weights, and reasoning data all open under Apache 2.0.

AA Intelligence Index v4.1: 17 (AA-estimated; well above the median for similar-size open-weight models) at a 262K-token context window. Versus K2-V2, hallucination rate drops from 89% to 52% and long-context reasoning improves from 33% to 53%. Tied for first place with K2-V2 and Olmo 3 on Artificial Analysis's Openness Index.