About K2

MBZUAI is releasing K2 with weights, code, training data, and checkpoints, delivering full-stack openness at a time when many “open” models reveal only fragments of their training process. K2 therefore stands out as one of the few high-performance models where every stage of development is documented, inspectable, and reproducible.

Created to advance the UAE’s leadership in open AI systems, K2 offers a strong foundation for assistants, scientific copilots, and agentic workflows that require both power and transparency. Its openness gives builders the visibility they need to create dependable, long-context, reasoning-centric applications.

AIME 2025

AIME 2025

K2 delivers strong reasoning performance relative to its size, outperforming larger open-weight models on AIME 2025. Its open training pipeline enables consistent gains across the Low, Medium, and High checkpoints, illustrating how transparent data and recipe design translate into competitive mathematical reasoning.

K2 benchmarked

Intelligence Index (40–150B Open Weights) Click to enlarge

According to data in the latest leaderboard published by artificialanalysis.ai, plotted across performance, openness, and model scale, the LLM landscape becomes unmistakably clear. Closed frontier U.S. models cluster on the far right, while Chinese open-weight systems fill the middle with partial transparency and heavy footprints. K2 stands alone: fully open, reproducible, and performant, offering a true open-source alternative in a space where no other viable option exists.