(MENAFN- Khaleej Times)

A UAE university has upgraded its artificial intelligence system to accelerate its race to compete with the world's best, and shared the complete blueprint for how they built it, something most tech companies refuse to do.

Mohamed bin Zayed University of Artificial Intelligence released the full version of K2, building on the smaller K2 Think model it launched earlier this year. This upgraded system can solve complex math problems, write computer code, and reason through difficult questions at the same level as OpenAI's latest models and China's top AI systems.

Recommended For You Why the future of media must be built on ethics, empathy, and shared responsibility

What makes this release unusual is its transparency. MBZUAI published the training instructions, the data they used, the development logs showing what worked and what didn't, and even the intermediate versions created during training. This step was made with the aim of setting the benchmark in open source models and their transparency.

"A model isn't truly open unless you also share the full training data, the logs showing what happened during development, and all the technical details," the university explained in its technical report.

This approach directly challenges current trends. Companies like OpenAI keep their development methods private. Even China's DeepSeek, which recently demonstrated that efficient design could rival billion-dollar systems, shared its final product but not the recipe behind it. The UAE is doing the complete opposite, offering both the system and the complete instructions for building it.

K2 performs as well as the world's leading AI systems on demanding tests. It correctly answers 69 per cent of graduate-level science questions and solves 83 per cent of complex logic puzzles - matching results from OpenAI's o3-mini and DeepSeek-R1. Among AI programs that anyone can freely access and modify, K2 leads in mathematical reasoning.

The system's development followed three stages. Researchers first taught it general knowledge, then specifically trained it on reasoning by feeding it 250 million math problems with step-by-step solutions, and finally fine-tuned it to understand instructions and use tools.

"We started by training K2-Think on carefully chosen examples of long, step-by-step reasoning," Hector Liu, who directs MBZUAI's Silicon Valley laboratory, told Khaleej Times earlier this year. "This helped the model learn to show its thought process, not just give answers."

Users can choose how much "thinking" they want K2 to do before answering. High-effort mode works best for difficult problems but takes longer. Low-effort mode handles everyday questions quickly. Researchers discovered that maximum thinking doesn't always produce better results-sometimes extended reasoning interferes with clear answers.

MBZUAI also openly addressed safety limitations. Testing across 72 scenarios showed K2 responds appropriately 86 per cent of the time. The researchers acknowledged a transparency challenge: reasoning AI might think through an unsafe answer but learn to mask those thoughts in its final response. By publishing this finding, they're inviting the global research community to help solve it.

K2 costs nothing to use for those who can run it on their own computers. Global Giant, OpenAI charges up to $4.40 per million words generated. DeepSeek charges less but still requires payment. K2 is completely free for researchers, students, and businesses willing to download and operate it themselves.

K2's update adds to a growing list of UAE-backed AI projects made fully accessible to researchers. By publishing the entire blueprint, MBZUAI says it hopes the model will serve as a reference point for how frontier systems can be built openly. More versions of K2 are expected in 2026 as the university expands its research into high-performance, transparent AI development.

MENAFN06122025000049011007ID1110446348