The Mohamed bin Zayed University of Artificial Intelligence, in partnership with Abu Dhabi-based technology group G42, has unveiled K2 Think, an open-source model that aims to put the UAE at the forefront of the artificial intelligence race.
The system, which is designed to tackle complex reasoning tasks, features 32 billion parameters. That makes it far smaller than some of the frontier AI models from OpenAI and DeepSeek, which often exceed 200 billion.
Yet MBZUAI said the system performs at a similar level on mathematical and scientific benchmarks, a breakthrough that could make high-end AI reasoning more efficient, affordable and widely accessible.
“Not only are we on par with them in terms of the performance of mathematical reasoning, we do it with a much smaller model, which means that you can get a much faster compute and also a much less expensive cost of generating the results,” Prof Eric Xing, president of MBZUAI, told The National.
The development signals the UAE’s growing ability to compete in the fast-evolving AI landscape. “We now have a home-grown technology to build a state-of-the-art system … as good as, if not better than, the best of its kind around the world, such as what’s coming out of OpenAI or DeepSeek,” he added.
In a social media post, Sheikh Tahnoon bin Zayed, Deputy Ruler of Abu Dhabi and National Security Adviser, who is also chairman of as Abu Dhabi holding company ADQ, said K2 Think was a “a significant step in advancing artificial intelligence from the UAE to the world”.
The launch reflects the UAE’s National AI Strategy 2031, which aims to position the country among the world’s leading AI powers.
While tech giants such as OpenAI and Anthropic have poured tens of billions into building ever-larger models, the race is increasingly shifting toward leaner systems that can deliver comparable reasoning power at a fraction of the cost.
Smarter, not bigger
What sets K2 Think apart is not scale, but efficiency. Internal testing showed it performed at the same level as the two best open-source reasoning models, DeepSeek V3.1 and OpenAI GPT-OSS, despite being much smaller, Prof Xing said.
The model's design is what sets it apart from the rest. "You don’t have to always rely on infrastructure and the capital to get the best reasoning outcome. You can be smarter,” Prof Xing explained. He added that better algorithms and architecture can deliver strong results more cost-effectively.
How it compares globally
K2 Think enters a field dominated by tech giants such as OpenAI, Anthropic and Google DeepMind, whose Gemini family of models is considered among the most advanced alongside OpenAI's GPT-4o and Anthropic's Claude 3.5.
MBZUAI said K2 Think narrows the gap in specialised reasoning, maths, coding and science, although the university does not claim that it matches the strongest systems across the board.
The distinction matters. While large language models have been popular for chat-style applications, the ability to reason through complex problems is viewed increasingly as the next frontier for AI, with potential applications in research, finance, logistics and engineering.
Applications and partnerships
Prof Xing said K2 Think could eventually serve as the backbone of business tools across industries including logistics and finance. Researchers at MBZUAI are already using the system to speed up mathematical derivations that might otherwise take weeks.
To use the model, MBZUAI worked with US-based Cerebras Systems, which builds wafer-scale processors. The specialised hardware allows K2 Think to run inference, the stage in which a model answers queries, up to 10 times faster than traditional GPU clusters.
G42, meanwhile, supported infrastructure development and will help to explore commercial use cases.
Such collaborations reflect a shift away from relying solely on Nvidia’s costly hardware. “You use 10,000 GPUs, I use 1,000, that’s cost-effectiveness,” Prof Xing said.
Open source and accountability
Another distinguishing feature is transparency. Unlike many leading AI systems, which are closed-source and reveal little about how they were trained, MBZUAI has made K2 Think’s data, training recipe and deployment code openly available.
The strongest models from companies such as OpenAI and Anthropic, including GPT-4o and Claude 3.5, are closed-source.
Users cannot download the weights, data or training recipe, those are only accessible through the system's own platforms or by using paid APIs. That means users pay to send queries to the company’s servers and receive answers without accessing the model itself.
Meta’s Llama models are more open, with weights for Llama 2 and Llama 3 released, but not the full training data or post-training recipe. That means they can be used, but not fully reproduced from scratch.
In contrast, MBZUAI said K2 Think can be replicated end-to-end, giving researchers the ability to study how reasoning emerges and adapt the model for new domains.
“Even criticisms and stress tests will help us improve the model further,” Prof Xing said. “We believe in growing with the community.”
Looking ahead
For Prof Xing, the real significance of K2 Think lies not only in the software itself, but in the institution behind it.
“K2 is the tip of the iceberg,” he said. “The real achievement is building an institution that can keep producing the talent, infrastructure and breakthroughs to solve the unsolved problems of tomorrow.”
MBZUAI plans to build on the K2 Think recipe for future models, including adaptations for health care and genomics. With efficiency as its hallmark, the university hopes the project will set a precedent – that the future of AI will not be defined only by size, but by smarter design.
MBZUAI, established in 2019, has been central to those efforts. In 2024, it released K2-65B, the world’s first fully reproducible open-source foundation model, a year after the launch of Jais, a large-scale Arabic language model.
