K2 Think 2nd Edition Unveiled By MBZUAI As Fully Sovereign Open-Source Reasoning Model

G42, Cerebras Systems and the Institute of Foundation Models at MBZUAI have released the second edition of the K2 Think reasoning model, a 70 billion parameter system based on the K2-V2 foundation. The updated model targets complex reasoning tasks and is positioned as MBZUAI’s flagship open-source general reasoning platform.

According to MBZUAI, this version of K2 Think represents the UAE’s first fully sovereign reasoning model, covering architecture, training and deployment. The model uses only Institute-curated data and offers transparency across pre-training, refinement, post-training alignment, evaluation and operational stages, aligning with national goals for technological sovereignty and responsible AI development.

K2 Think is now described as a 70 billion parameter open-source general reasoning model built on K2-V2 Instruct, which MBZUAI calls its most capable fully open-source instruct system. The model shows stronger performance on demanding reasoning benchmarks, including AIME2025, HMMT, GPQA-Diamond and IFBench, when compared with other available open-source reasoning systems.

The university reports that K2 Think has been engineered for step-by-step, long-form reasoning across mathematics, science, programming and logical problem-solving. These capabilities aim to approach the performance frontier in domains such as mathematics, coding and STEM, with data mainly drawn from these areas, then carefully deduplicated from K2-V2 Instruct training data and decontaminated from downstream evaluation sets.

Following the first K2 Think release in September 2025 and K2-V2 in December 2025, MBZUAI refined its Guru dataset. Updates included extending coverage to more domains, adding difficulty-based filtering informed by K2-V2, and enforcing decontamination against major evaluation benchmarks to reduce data leakage and improve scientific reliability.

The STEM section of the Guru dataset was enlarged using science questions from the Nemotron post-training dataset that were excluded from K2-V2 instruct tuning. MBZUAI states that every stage of K2 Think’s development is open to independent review and reproduction, which supports reproducible AI research and keeps the model globally accessible while remaining sovereign.

MBZUAI notes that with this release, its K2 Think reasoning model is fully sovereign and open from pre-training through post-training. The institute expects these open-source efforts, combined with performance gains, to narrow the gap between community-owned reproducible AI models and proprietary systems, while supporting the UAE’s broader AI innovation strategy.

With inputs from WAM