MBZUAI Unveils K2 Think V2, Advancing the UAE’s Push for Fully Sovereign AI
The 70B-parameter system embeds reasoning directly into a fully open foundation model.
News
- MBZUAI Unveils K2 Think V2, Advancing the UAE’s Push for Fully Sovereign AI
- ENEC, TII, and ASPIRE Partner to Advance Autonomous Aerial Systems
- OpenAI Bets on AI to Streamline Scientific Research Workflows
- AI Use Is Rising at Work but Productivity Gains Lag
- Meta Signals Shift Toward Paid Features With Subscription Tests Across Its Apps
- All LLMs Are Trained on the Same Data’: Oracle Chief Flags AI’s Biggest Challenge
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), alongside G42 and Cerebras Systems, has launched K2 Think V2, a 70-billion-parameter reasoning system, positioning it as a significant step toward the UAE’s ambition to build an end-to-end sovereign AI infrastructure.
Developed by MBZUAI’s Institute of Foundation Models (IFM), K2 Think V2 is built on the newly released K2-V2 base model, which the university describes as its strongest frontier-class, open foundation model to date. The release marks a shift from earlier iterations of K2 Think, which were open at the model level but not across the full lifecycle. K2 Think V2 is presented as open “from end to end,” spanning data curation, pre-training, post-training, alignment, and evaluation.
This level of transparency is central to the project’s claim of technological sovereignty. According to MBZUAI, all training data is curated in-house, including the Guru dataset, and is decontaminated from downstream benchmarks to reduce evaluation bias. Intermediate checkpoints, training recipes, and evaluation results are also made publicly available, allowing independent reproduction and scrutiny—an approach that contrasts with the partial openness typical of many large-scale AI systems.
Technically, K2 Think V2 is designed as a reasoning-first system, with reasoning capabilities embedded directly into the foundation model rather than layered on top. The model supports extended context lengths, enabling sustained multi-step reasoning across mathematics, science, coding, logic, and simulation.
MBZUAI reports that this architecture delivers strong performance on demanding reasoning benchmarks, including AIME 2025, GPQA-Diamond, HMMT, and IFBench, compared with other open-source reasoning models.
The launch builds on MBZUAI’s and its partners’ broader strategy to position the UAE as a producer, not merely a consumer, of foundational AI systems. In December 2025, the same consortium introduced Jais 2, an open-weight Arabic large language model trained on what they described as the richest Arabic-first dataset to date.
