MBZUAI launches K2 Think V2 as a new artificial intelligence reasoning system developed in the United Arab Emirates. The model was introduced by the Mohamed bin Zayed University of Artificial Intelligence. It is built on the K2-V2 foundation model, which contains 70 billion parameters. The system is described as fully sovereign and openly developed from training to evaluation.
Model Design and Development
K2 Think V2 integrates reasoning directly into the foundation model architecture. It does not rely on external reasoning layers. The design supports long-context processing and structured reasoning. MBZUAI states that the full development process is transparent and reproducible. This includes dataset selection, training methods, and testing procedures.
The base K2-V2 model was created by MBZUAI’s Institute of Foundation Models. It is designed to support advanced reasoning workloads across multiple domains.
Reasoning Tasks and Benchmarks
The model is designed to perform reasoning tasks in mathematics, science, coding, logic, and simulation. K2 Think V2 has been evaluated using recognized benchmarks. These include AIME2025, GPQA-Diamond, HMMT, and IFBench. The reported results place the model among competitive open reasoning systems.
The system supports extended chains of reasoning and multi-step problem solving.
Access and Related AI Developments
MBZUAI launches K2 Think V2 with support from partners including G42 and Cerebras Systems. The model is available through web and mobile platforms. Users can access it with voice input and without paid subscriptions.
The same AI news update also references Stanford University’s Center for Research on Foundation Models. Stanford announced a new Arabic large language model benchmarking leaderboard developed with Arabic.AI. The leaderboard focuses on standardized evaluation of Arabic-language models.
K2 Think V2 follows earlier K2 model releases and represents the latest phase of MBZUAI’s foundation model research.
Source: https://www.middleeastainews.com/p/mbzuai-launches-k2-think-v2-stanford
