Stay informed with free updates
Simply sign up to the Artificial intelligence myFT Digest — delivered directly to your inbox.
Abu Dhabi’s top tech university has launched an AI system capable of advanced reasoning that researchers say ranks alongside the best open models from the US and China, accelerating the United Arab Emirates’ push into “sovereign” AI.
The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) on Tuesday released its latest model, K2 Think, alongside a complete disclosure of the data, algorithms and other code that was used to build it, as the country seeks to gain an edge in a market that has become dominated by Chinese AI groups.
Eric Xing, MBZUAI president, said the state-backed university was filling a gap left by Silicon Valley firms such as Meta that have pulled back from openly publishing the research behind their AI systems.
“In the western community there hasn’t been an answer to the Chinese open-weight models yet. Our production is filling that void,” he told the FT. “We want to provide safe and reliable options for start-ups and other researchers to carry out additional work.”
It comes as recent research showed that Chinese groups such as DeepSeek and Alibaba had overtaken their US rivals in the global market for open AI models for the first time last year.
MBZUAI president Eric Xing: ‘In the western community there hasn’t been an answer to the Chinese open-weight models yet. Our production is filling that void’ © Abaca Press/Alamy
Abu Dhabi is investing heavily in AI infrastructure, including plans for a huge data-centre cluster for the OpenAI-led “Stargate” project. The emirate’s state-backed AI fund MGX has also invested in OpenAI as well as striking infrastructure partnerships with US firms including Microsoft, Silver Lake and BlackRock.
The Gulf state’s ability to use homegrown resources — powered by US-based cloud computing services — to build a globally competitive AI system comes as many countries are reliant on the US and China for a technology that could improve public services and upend their economies.
“There is a non-trivial risk factor for using even open-weight models that you don’t know how they were produced,” Xing said.
MBZUAI’s previous release of K2 was criticised by researchers at ETH Zurich for exaggerating its capabilities but a study by Artificial Analysis, which tests and ranks AI systems, found the latest version performed at a similar level to equivalent open models from Nvidia, OpenAI and Alibaba.
It scored particularly well on reducing the number of errors or “hallucinations” that can make the technology too unreliable for many applications.
“It is the most intelligent model at this level of openness [and] it contributes to the status of the United Arab Emirates as an emerging contender in the AI race,” said Micah Hill-Smith, co-founder and chief executive of Artificial Analysis. He noted that it tied for leadership in its “Openness index”, which ranks models based on their transparency.
MBZUAI’s K2 Think was trained for a fraction of the cost of OpenAI, Google or Anthropic’s latest models © MBZUAI
Artificial Analysis is an independent firm that tests every major new AI system. It was paid a fee by MBZUAI to assess K2 Think ahead of its public launch.
Keegan McBride, senior policy adviser on emerging technology and geopolitics at the Tony Blair Institute, said that K2’s transparency had won its developers respect in the open-source AI community.
“The UAE is investing a lot of resources and they continue to punch above their weight,” he said. “Really what we are seeing is a country that has taken AI seriously for a long time reaping the benefits of that now.”
K2 Think was trained for a fraction of the cost of OpenAI, Google or Anthropic’s latest models and used fewer than 2,000 of Nvidia’s H200 chips.
The improvements to K2 Think’s capabilities were achieved by replacing the off-the-shelf open-weight model used to train its predecessor with a new base model built fully in-house, Xing said. “K2 Think is top to bottom homemade,” he added. “Because we publish the data and the training recipe, there is nothing we can hide.”
Jasper Dekoninck, a PhD student at ETH Zurich, noted that MBZUAI had done “decontamination checks” to avoid using training data that could skew benchmarks, addressing some of his previous concerns.
But he added that K2 Think’s claim of being the “best fully open-source model” was a less significant milestone because so few other models met the same transparency criteria.