The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi has unveiled K2 Think, a low-cost reasoning model designed to rival systems developed by DeepSeek and OpenAI.
The launch underscores the UAE’s ambitions to become a significant player in the global AI sector, currently dominated by the United States and China.
A new class of reasoning model
MBZUAI describes K2 Think as “a new class of reasoning model.” Instead of relying on massive parameter counts, the system emphasizes efficient architecture and logical depth. According to the university, it uses long chain-of-thought supervised fine-tuning to improve reasoning and reinforcement learning with verifiable rewards to enhance accuracy on complex problems.
Hector Liu, director of MBZUAI’s institute of foundation models, said the project is treated “more like a system than just a model.” Unlike typical open-source releases, K2 Think is being actively deployed and iteratively improved, reflecting a long-term development strategy.
Speed and efficiency
K2 Think is built on Alibaba’s Qwen 2.5 large language model and runs on hardware from Cerebras. MBZUAI claims it is among the fastest and most efficient reasoning systems, processing around 2,000 tokens (about 1,500 words) per second. This balance of speed and accuracy is intended to give it a competitive edge against larger, more resource-intensive models.
Open-source and transparent
Like DeepSeek’s R1 model, K2 Think is open source. Its training data and weights are publicly available, allowing researchers worldwide to study, reproduce, and extend its reasoning methods. MBZUAI says this transparency will encourage collaboration and accelerate progress in reasoning models.
Strategic importance for the UAE
The university calls K2 Think “a defining moment for AI in the UAE,” emphasizing how open innovation and public–private partnerships can position Abu Dhabi as a global AI hub.
By investing in efficient and transparent AI systems, the UAE aims to influence how reasoning models evolve beyond scale-driven approaches.
