The United Arab Emirates has added to the global artificial intelligence race with its K2 Think reasoning model, which it hopes can compete with flagship models from developers such as Microsoft-backed (MSFT) OpenAI and China’s DeepSeek (DEEPSEEK) despite being only a fraction of the size.
The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence and G42, an AI development holding company based in Abu Dhabi, announced the release of K2 Think on Tuesday.
It was built with 32B parameters but outperforms flagship models that are more than 20 times its size, according to its creators. For example, OpenAI’s GPT-3.5 model has 175B parameters, while GPT-4 is estimated to have more than 1 trillion. DeepSeek-R1 has 671B parameters, but only 37B parameters are activated during a response to a given input.
K2 Think was built with Alibaba’s (NYSE:BABA) open-source Qwen 2.5 model. MBZUAI said it will soon be available on Cerebras Systems’ computing hardware. Cerebras designs AI chips and systems and is one of the few burgeoning competitors to the GPU giant Nvidia (NVDA). Cerebras is attempting to launch an IPO but has faced scrutiny from U.S. regulators due to national security questions over some of its investors.
K2 Think is fully open source, allowing global researchers full transparency into its training data and parameter weights.
“K2 Think, developed by MBZUAI’s Institute of Foundation Models, is a significant advancement for the global AI research and development community,” said MBZUAI president and professor Eric Xing. “By delivering these advances in a fully transparent framework, we are ushering in a new era of cost-effective, reproducible and accountable AI.”