emirates7 - G42, Cerebras Systems and the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) on Tuesday launched the second edition of the K2 Think model.
The new advanced reasoning system, which includes 70 billion parameters, is built on the K2-V2 base model, the Institute’s strongest frontier-class open-source foundation model, designed specifically to support K2 Think.
MBZUAI said the second edition marks a major milestone in the UAE’s technological sovereignty journey, as it represents the first fully sovereign reasoning model in terms of architecture and operation, providing full transparency across pre-training data, refinement mechanisms, post-training alignment, evaluation and deployment.
The university explained that the model has been designed to solve complex problems step by step through long-form reasoning across areas such as mathematics, science, programming and logic, delivering leading results on advanced reasoning benchmarks compared to other open-source systems.
MBZUAI added that all stages of the model’s development are available for independent inspection and reproduction, supporting scientific reliability and reinforcing national objectives for sovereign AI innovation while maintaining global accessibility.
K2 Think is now a 70 billion parameter open-source general reasoning model, built on top of K2-V2 Instruct, the most capable fully open-source instruct model. It delivers improved performance across complex reasoning benchmarks, including AIME2025, HMMT, GPQA-Diamond and IFBench.
With this release, MBZUAI’s flagship reasoning model is now fully sovereign and open from pre-training through post-training, using only Institute-curated data. These open-source efforts, alongside improved performance, are helping close the gap between reproducible AI owned by the global community and proprietary models.
Following the introductory release of K2 Think in September 2025 and the release of K2-V2 in December 2025, MBZUAI made several improvements to its existing Guru dataset, including expanding to more domains, applying difficulty-based filtering based on K2-V2, and ensuring decontamination against key evaluation benchmarks.
The STEM portion of Guru was expanded by incorporating science questions from the Nemotron post-training dataset that were not used during the instruct tuning phase of K2-V2.
MBZUAI said the strategy of the original K2 Think model remains focused on approaching the frontier of key domains such as mathematics, coding and STEM. Data were selected mainly from these domains, carefully deduplicated from the training data used for K2-V2 Instruct and fully decontaminated from downstream evaluations.
The new advanced reasoning system, which includes 70 billion parameters, is built on the K2-V2 base model, the Institute’s strongest frontier-class open-source foundation model, designed specifically to support K2 Think.
MBZUAI said the second edition marks a major milestone in the UAE’s technological sovereignty journey, as it represents the first fully sovereign reasoning model in terms of architecture and operation, providing full transparency across pre-training data, refinement mechanisms, post-training alignment, evaluation and deployment.
The university explained that the model has been designed to solve complex problems step by step through long-form reasoning across areas such as mathematics, science, programming and logic, delivering leading results on advanced reasoning benchmarks compared to other open-source systems.
MBZUAI added that all stages of the model’s development are available for independent inspection and reproduction, supporting scientific reliability and reinforcing national objectives for sovereign AI innovation while maintaining global accessibility.
K2 Think is now a 70 billion parameter open-source general reasoning model, built on top of K2-V2 Instruct, the most capable fully open-source instruct model. It delivers improved performance across complex reasoning benchmarks, including AIME2025, HMMT, GPQA-Diamond and IFBench.
With this release, MBZUAI’s flagship reasoning model is now fully sovereign and open from pre-training through post-training, using only Institute-curated data. These open-source efforts, alongside improved performance, are helping close the gap between reproducible AI owned by the global community and proprietary models.
Following the introductory release of K2 Think in September 2025 and the release of K2-V2 in December 2025, MBZUAI made several improvements to its existing Guru dataset, including expanding to more domains, applying difficulty-based filtering based on K2-V2, and ensuring decontamination against key evaluation benchmarks.
The STEM portion of Guru was expanded by incorporating science questions from the Nemotron post-training dataset that were not used during the instruct tuning phase of K2-V2.
MBZUAI said the strategy of the original K2 Think model remains focused on approaching the frontier of key domains such as mathematics, coding and STEM. Data were selected mainly from these domains, carefully deduplicated from the training data used for K2-V2 Instruct and fully decontaminated from downstream evaluations.