Omer Taha Cetin |Anadoru | Getty Images
A new challenger in the global artificial intelligence race has entered the ring.
Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), an AI-centric research university founded by the United Arab Emirates, announced on Tuesday the release of a new, low-cost inference model comparable to Openai and Deepseek.
Earlier this year, Chinese AI lab Deepseek shocked the world with the release of an inference model called the R1, and said it could surpass Openai, but it would cost much less training.
With just 32 billion parameters, Mbzuai’s model, called K2 Think, is much smaller than the Openai and Deepseek competitive systems. It was built on top of it Alibaba’s The open source Qwen 2.5 model is run and tested on hardware provided by AI chip maker Cerebus.
In the context, DeepSeek’s R1 has a total of 671 billion parameters. This is basically another term for variables that AI language models learn to understand and generate languages. Openai does not disclose parameter counts for AI models.
K2 thought it was developed in collaboration with G42, a Buzzy UAE-based AI company supported by US Tech Giant Microsoft. The researchers behind it say it offers performance comparable to Openai and Deepseek’s flagship inference models, despite being only a small part of its size.
They cited the benchmarks AIME24, AIME25, HMMT25, and Omni-Math-Hard related to mathematics, benchmark livecodebenchv5, and science benchmark GPQA-diamonds.
How did they do that?
Hector Liu, director of Mbzuai’s Foundation Model Institute, told CNBC that the team behind K2 believes they can achieve such high-level performance using many methods.
They include long chain chains (COT) monitored fine-tuning (step-by-step method of inference) and test time scaling to improve performance by allocating additional computing resources during “inference” or to apply learning knowledge to data that has never been seen before.
“What’s special about our models is that we treat them like systems rather than just models,” Liu told CNBC. “So unlike the usual open source models that can release models, we’ll actually deploy them and see how to improve them over time.”
“It’s very difficult to say when you ask which of the single steps is most important. It’s like a systematic way of combining all of these methods,” he added.
Why is it important?
On the world stage there are two countries that stand out as pioneers of AI racing: the United States and China.
American tech giants and startups like Openai led early momentum with what is known as basic models. But Deepseek’s breakthrough with R1 earlier this year strengthened China’s position as a formidable AI player in itself.
More recently, the United Arab Emirates has sought to establish itself as a global leader in AI to increase geopolitical influence and diversify the economy beyond crude oil dependence.
This region can be used to refer to AI development company G42 as an example of how it has achieved its position within the space. However, it faces fierce competition with nearby Saudi Arabia, which is seeking to develop full-stack AI capabilities through Humain, a company launched under the public investment fund in May.
Beyond that, there is also the geopolitical complexity that surrounds the ambitions of the UAE AI. Microsoft’s investment and partnership with the G42 last year received much scrutiny in the US in relation to its relationship with China.
More broadly, the UAE AI industry has come a long way to reach the size of its US and Chinese counterparts. Openai and the big tech players enjoy a good head start with their respective basic AI models, but Beijing has long viewed AI as a strategic priority.
Focus on scientific breakthroughs
Although K2 believes it demonstrates performance on par with Openai, the system developer says its purpose is not to build a chatbot like ChatGpt. Richard Morton, managing director of Mbzuai’s Foundation Model Research Institute, explains that the model is intended to provide specific uses in fields such as mathematics and science.
“The fact is that the basic reasoning of the human brain is the cornerstone of all thought processes,” Morton told CNBC.
“This particular application is very condensing this period, like taking 1,000 or 2,000 people for five years to think about a particular question, but going through a specific clinical trial or something.”
It also expands the scope of advanced AI technologies in regions where capital and infrastructure are not accessible by US companies.
“What we’re discovering is that the less you can do,” Morton said.
