The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42, based in Abu Dhabi, have unveiled K2 Think, an open-source AI reasoning system that matches the performance of the top reasoning models while being significantly smaller in size.
K2 Think boasts 32 billion parameters and outshines leading reasoning models that are nearly 20 times larger, as stated by MBZUAI’s Institute of Foundation Models.
This system ranks among the leading reasoning models worldwide, excelling in various mathematical benchmarks, such as AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD.
Features of K2 Think
K2 Think is founded on six key innovation principles, utilizing chain-of-thought supervised fine-tuning, reinforcement learning with verifiable rewards, agent planning, and scalable operations during testing.
The model will also be accessible on Cerebras’ wafer-scale computing platform, optimized for processing at 2,000 tokens per second.
Khaldoon Khalifa Al Mubarak, chairman of MBZUAI’s Board of Trustees and member of the Artificial Intelligence and Advanced Technology Council (AIATC), remarked: “The global standard established by K2 Think reflects the groundbreaking work of MBZUAI’s Institute of Foundation Models initiative, showcasing the UAE’s dedication to creating sophisticated systems developed locally and shared globally.”
Peng Xiao, a board member of MBZUAI and group CEO of G42, noted: “K2 Think has transformed the AI reasoning landscape from focusing on size to prioritizing intelligence. This milestone demonstrates that more compact, efficient models can compete with the largest systems, heralding a new era of AI advancements.”
Differing from many models that provide only weights, K2 Think is entirely open source, offering training data, weights, and deployment software.
Professor Eric Xing, president of MBZUAI, stated that the system symbolizes “a new phase of cost-efficient, reproducible, and accountable AI.”
MBZUAI’s Open-Source Model Collection
K2 Think enhances MBZUAI’s collection of open-source models, which includes Jais (Arabic), NANDA (Hindi), and SHERKALA (Kazakh), while also building on the K2-65B, the first fully reproducible open-source foundation model released earlier this year.
K2 Think is accessible at K2Think.Ai/K2Think as well as on Hugging Face.