The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in the UAE has enhanced its artificial intelligence system to boost its global competitive edge, sharing the entire blueprint that most tech companies keep secret.
MBZUAI has unveiled the full K2 model, expanding on the smaller K2 Think model released earlier this year. This upgraded system matches the capabilities of top AI models like OpenAI’s and China’s, solving complex math issues, writing code, and answering challenging questions.
This release stands out for its transparency. MBZUAI has shared the training guidelines, data utilized, and development logs, setting a new standard for open-source models.
Stay updated with the latest news. Follow KT on WhatsApp Channels.
“A model isn’t truly open unless all technical details, including full training data and development logs, are shared,” the university stated in its report.
This approach contrasts with current industry norms. Companies like OpenAI keep their processes private, and even China’s DeepSeek, despite demonstrating efficient design, didn’t disclose its development process. The UAE, however, shares both the system and its complete building instructions.
K2 performs equally well on demanding AI evaluations, correctly answering 69% of graduate-level science questions and solving 83% of complex logic puzzles, comparable to OpenAI’s o3-mini and DeepSeek-R1. Among open-access AI programs, K2 excels in mathematical reasoning.
The system’s development included three phases: initial general knowledge training, focused reasoning training with 250 million math problems, and final fine-tuning for tool use and instructions.
“We trained K2-Think using examples of detailed, step-by-step reasoning,” said Hector Liu from MBZUAI’s Silicon Valley lab to Khaleej Times. “This approach helped the model articulate its thought process rather than just providing answers.”
Users can adjust K2’s “thinking” level before answering. High-effort mode is ideal for complex queries but takes longer, while low-effort mode is faster for everyday questions. Researchers found that more thinking doesn’t always enhance accuracy.
MBZUAI also addressed safety concerns, noting that K2 responded appropriately 86% of the time in 72 scenarios. They highlighted a challenge: reasoning AI may consider unsafe responses but mask them in final answers. They invite global researchers to tackle this issue.
K2 is free for anyone capable of running it on their computers, contrasting with OpenAI’s $4.40 per million words and DeepSeek’s fees. Researchers, students, and businesses can download and operate it without cost.
K2’s release adds to a series of UAE-supported AI initiatives available to researchers. By sharing the full blueprint, MBZUAI aims for the model to be a benchmark for transparent AI development. More K2 versions are expected in 2026, as the university continues to advance high-performance, transparent AI research.
