Researchers in Abu Dhabi say K2 Think matches reasoning models from OpenAI and DeepSeek while using far fewer resources.
The model was built at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi and is being released as open source by G42, an Emirati technology conglomerate backed by Abu Dhabi’s sovereign wealth funds. G42 is operating the system on a cluster of Cerebras chips, a hardware approach that differs from Nvidia’s dominant GPUs.
Officials present K2 Think as a clear signal that the United Arab Emirates’ heavy investments in artificial intelligence are beginning to yield results. The project is one of the country’s entries in an international competition to show technical strength in a field with major economic and geopolitical stakes. The United States and China remain the two largest competitors, and a number of well-funded smaller states are racing to field their own sovereign AI systems.
K2 Think is compact by current industry measures, at about 32 billion parameters. Its designers did not aim to build a general-purpose large language model. Instead they focused on reasoning: training the system to work through complicated questions by simulating chains of thought and stepwise deliberation rather than by quickly producing a single synthesized answer. For those kinds of reasoning challenges, MBZUAI researchers say K2 Think’s performance is on par with reasoning engines from OpenAI and DeepSeek, both of which use models with more than 200 billion parameters.
“This is a technical innovation or, in my opinion, a disruption,” Eric Xing, MBZUAI’s president and lead AI researcher, said ahead of the announcement.
Xing credited the result to a careful mix of recent technical methods. The team fine-tuned the model on long sequences that mimic explicit, step-by-step reasoning. They added an agentic planning layer that decomposes problems into multiple alternative paths, and they used reinforcement learning to favor answers that can be checked for correctness. Engineering tweaks let the system run efficiently on Cerebras hardware so that inference and training can be delivered with lower overhead than would be expected for larger architectures.
“How to make a smaller model function as well as a more powerful one—that’s a lesson to learn, if other people want to learn from us,” Xing said.
He told reporters that the project’s early phases relied on several thousands of GPUs, though he declined to provide a precise total. The final training pass used something on the order of 200 to 300 chips. The research team plans to fold K2 Think into a complete large language model in the coming months. MBZUAI has published the model’s weights and code under an open license and released a technical report that lays out the training recipes, architectural choices, and trade-offs the group used to reach its results.
Across the Middle East, governments and private groups are pouring money into AI labs, data centers, and research networks. Saudi Arabia has been one of the region’s largest investors in infrastructure and talent recruitment tied to advanced computing. In May, President Donald Trump visited the region and announced a slate of AI-related deals that involved multiple U.S. technology firms.
Peng Xiao, CEO of G42, and a MBZUAI board member, said in a statement: “By proving that smaller, more resourceful models can rival the largest systems, this achievement shows how Abu Dhabi is shaping the next wave of global innovation.”