Omer Taha Cetin | Anadolu | Getty Images
A new challenger in the global artificial intelligence race has entered the ring.
The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), an AI-focused research university established by the United Arab Emirates, announced on Tuesday the release of a new, low-cost reasoning model to rival OpenAI and DeepSeek.
It comes after DeepSeek, a Chinese AI lab, earlier this year shocked the world with the release of a reasoning model called R1 which it said could outperform OpenAI but with far less training costs.
At just 32 billion parameters, MBZUAI’s model, dubbed K2 Think, is much smaller than competing systems from OpenAI and DeepSeek. It was built on top of Alibaba’s open-source Qwen 2.5 model and is run and tested on hardware provided by AI chipmaker Cerebas.
For context, DeepSeek’s R1 has a total of 671 billion parameters, which is essentially another term for the variables that an AI language model learns to understand and generate language. OpenAI doesn’t disclose the parameter counts of its AI models.
K2 Think was developed in partnership with G42, the buzzy UAE-based AI firm backed by U.S. tech giant Microsoft. The researchers behind it say it delivers performance on par with the flagship reasoning models of OpenAI and DeepSeek — despite being a fraction of the size.
They cited the benchmarks AIME24, AIME25, HMMT25 and OMNI-Math-HARD, which relate to math, coding benchmark LiveCodeBenchv5 and science benchmark GPQA-Diamond.
How did they do it?
Hector Liu, director of MBZUAI’s Institute of Foundation Models, told CNBC the team behind K2 Think were able to achieve such high levels of performance by using a number of methods.
They include long chain-of-thought (CoT) supervised fine-tuning — a method of step-by-step reasoning — as well as so-called test-time scaling, which is a technique for improving performance by allocating extra computing resources during “inferencing” — or, applying learned knowledge to data it’s never seen before.
“What was special about our model is we treat it more like a system than just a model,” Liu told CNBC. “So, unlike a regular open-source model where we can just release the model, we actually deploy the model and see how we can improve the model over time.”
“If you ask me which one of the single steps is the most important, it’s very hard to say. It’s more like a system method work where all these methods combined delivered the final result,” he added.
Why does it matter?
There are two countries on the world stage that stand out as the forerunners in the AI race: the U.S. and China.
America’s tech giants and startups like OpenAI led the early momentum with so-called foundation models, which aim to fulfill a wide range of tasks by relying on vast amounts of training data. However, DeepSeek’s breakthrough with R1 earlier this year reinforced China’s position as a formidable AI player in its own right.
More recently, the UAE has sought to position itself as a global leader in AI in a bid to enhance its geopolitical influence and diversify its economy beyond crude oil dependency.
The region can point to its AI development firm G42 as an example of how it’s gaining ground in the space. However, it faces fierce competition from neighboring Saudi Arabia, which is looking to develop full-stack AI capabilities via Humain, a company launched under the Public Investment Fund in May.
Beyond that, there are also geopolitical complexities that shroud the UAE’s AI ambitions. Microsoft’s investment and partnership with G42 last year attracted a great deal of scrutiny in the U.S. related to the company’s relationship with China.
More broadly, the UAE’s AI industry still has a long way to go to reach the scale of its U.S. and Chinese counterparts. OpenAI and the Big Tech players have enjoyed a good head start with their respective foundation AI models, while Beijing has long considered AI a strategic priority.
Focus on scientific breakthroughs
While K2 Think demonstrates performance on par with OpenAI, the system’s developers say the aim is not to build a chatbot like ChatGPT. Richard Morton, managing director for MBZUAI’s Institute of Foundation Models, explains the model is intended to serve specific uses in fields like math and science.
“The fact is that the fundamental reasoning of the human brain is the cornerstone of all the thinking process,” Morton told CNBC.
“With this particular application, instead of taking 1,000, 2,000 human beings five years to think through a particular question, or go through a particular set of clinical trials or something like that, this vastly condenses that period.”
It could also expand the reach of advanced AI technologies in regions that don’t have access to the kind of capital and infrastructure U.S. firms possess.
“What we’re discovering is that you can do a lot more with less,” Morton said.
