This approach is intended to deliver a system that is both more transparent and more capable, while strengthening independent AI development capacity within the country.
G42, Cerebras Systems and the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have announced the release of K2 Think V2, a new 70-billion parameter advanced reasoning system built on the K2-V2 base model. The model represents IFM’s most powerful frontier-class open-source foundation model to date and has been purpose-built to power the K2 Think reasoning architecture.
The launch is being positioned as a major step forward for the UAE’s technological sovereignty. While earlier versions of K2 Think were openly accessible, K2 Think V2 is the first iteration to be open end-to-end, covering the full lifecycle from pre-training data and curation to post-training processes, reasoning alignment and evaluation. This approach is intended to deliver a system that is both more transparent and more capable, while strengthening independent AI development capacity within the country.
By upgrading to the K2-V2 foundation model, K2 Think V2 introduces higher levels of performance, openness and autonomy. The development reinforces the UAE’s strategy of building frontier-grade AI systems that are globally accessible while remaining fully sovereign. The shift also reflects a broader move from layering reasoning capabilities on top of an existing foundation model to embedding reasoning directly into the core model architecture.
K2 Think V2 inherits long-context capabilities and full training transparency from K2-V2, allowing it to function as a fully sovereign system across the entire stack. Every stage of the model’s development is open, inspectable and independently reproducible, supporting scientific credibility as well as national AI independence. The system is designed specifically for step-by-step reasoning, enabling it to tackle complex problems across mathematics, science, coding, logic and simulation through extended chains of thought.
This reasoning-first and fully open design translates into stronger performance on demanding benchmarks. K2 Think V2 has reported leading results among open-source reasoning systems on evaluations such as AIME2025, GPQA-Diamond, HMMT and IFBench, highlighting its capabilities in advanced problem-solving tasks.
The model is built on IFM’s latest foundation architecture, which was designed from the outset to support reasoning, long context handling and alignment. Expanded context length enables the system to sustain multi-step reasoning over significantly larger volumes of information, supporting deeper analytical processes.
Training for K2 Think V2 relies entirely on IFM-curated datasets, including the Guru dataset, which has been fully decontaminated from downstream benchmarks to ensure fair and reliable evaluation. The project also emphasizes what it describes as “360-open” transparency, with pre-training data, intermediate checkpoints, post-training methods and evaluation materials made available for inspection, reuse and further development by the research community.
