MBZUAI launches K2 Think, a low-cost reasoning model to rival global AI giants

The model is built on Alibaba's open-source Qwen 2.5 and tested using hardware from AI chipmaker Cerebras. It was developed in collaboration with G42, a UAE-based firm backed by Microsoft.

By  Storyboard18| Sep 11, 2025 9:56 AM
While DeepSeek's flagship R1 model boasts 671 billion parameters, K2 Think is significantly smaller with just 32 billion parameters, according to media reports.

The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the UAE has unveiled K2 Think, a low-cost reasoning model designed to compete with major players like OpenAI and China’s DeepSeek.

While DeepSeek's flagship R1 model boasts 671 billion parameters, K2 Think is significantly smaller with just 32 billion parameters, according to media reports. The model is built on Alibaba's open-source Qwen 2.5 and tested using hardware from AI chipmaker Cerebras. It was developed in collaboration with G42, a UAE-based firm backed by Microsoft, stated media reports.

Despite its smaller size, K2 Think performs competitively on several key benchmarks, including:

???? AIME24, AIME25

???? OMNI-Math-HARD, GPQA-Diamond (science reasoning)

???? LiveCodeBenchv5 (coding)

???? HMMT25

According to a Moneycontrol report, Hector Liu, head of MBZUAI’s Institute of Foundation Models, noted that the model’s performance was achieved by combining techniques like chain-of-thought supervised fine-tuning for step-by-step reasoning, and test-time scaling, which enhances inference by applying extra compute during runtime.

First Published onSep 11, 2025 9:56 AM

The Grand Irony: Agencies That Built Brands, Forgot to Build For Themselves

Despite being the original architects of global brands, advertising holding companies are collapsing in market value because they still sell human hours while the world now rewards scalable, self-learning systems.