Abu Dhabi Challenges DeepSeek & OpenAI with 'K2 Think,' A Low-Cost AI Reasoning Model

On Tuesday, September 9, the United Arab Emirates (UAE) introduced a low-cost AI reasoning model as a substitute for comparable products from companies like DeepSeek and OpenAI.

Interestingly, the UAE-established Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), a research university with an emphasis on AI, collaborated with Microsoft-backed local AI company G42 to create K2 Think.

K2 Think: The New AI Model

A report indicates that it was created with the open-source Qwen 2.5 model from the massive Chinese internet company Alibaba and is driven by the AI chip manufacturer Cerebas.

Remarkably, DeepSeek's R1 model includes 671 billion parameters, several times more than the newly revealed AI model, which has 32 billion parameters total. A large language model's (LLM) ability to learn and provide sophisticated replies is reflected in the number of parameters it has.

Meanwhile, the UAE hopes to establish itself as a contender in the global competition for supremacy in AI, which has up to now been mostly presented as a US-China struggle, with the introduction of K2 Think. India and other nations are stepping up their attempts to attain AI sovereignty at the same time.

Notably, Liu, director of MBZUAI's Institute of Foundation Models, defined their model's distinguishing feature as being regarded more like a system than a model. This technique varies from traditional open-source models, which are merely released; instead, the model is deployed to allow for ongoing refinement over time.

According to reports, K2 Think is intended to facilitate scientific discoveries. The truth is that the foundation of all thought processes is the basic reasoning of the human brain.

Richard Morton, managing director of MBZUAI's Institute of Foundation Models, was quoted as saying, “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.”

Performance & Model Specifications

K2 Think has been further tuned to attain higher levels of performance with fewer computer resources, even if its source code and core technology appear to be built on Alibaba's Qwen 2.5 open-source AI model.

The paper cites several methods, including test-time scaling and chain-of-thought (CoT) supervised fine-tuning, to achieve this.

Furthermore, various benchmarks relating to maths, coding, and science, including AIME24, AIME25, HMMT25, OMNI-Math-HARD, LiveCodeBenchv5, and GPQA-Diamond, were used to assess the model.

According to the researchers who developed K2 Think, the model's performance was comparable to that of state-of-the-art AI reasoning models from OpenAI and DeepSeek while being much smaller.

Read more news:

In a Major Milestone, London Premier Centre Achieves PMI Recognition

LPC Training Achieves ISO 9001:2015 Accreditation from British Assessment Bureau

LPC Celebrating CMI Recognition