The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), in collaboration with G42, a company based in Abu Dhabi, has introduced K2 Think. This open-source system is designed for advanced AI reasoning, offering capabilities comparable to the most robust reasoning models while occupying significantly less space.
K2 Think boasts 32 billion parameters, achieving results superior to leading reasoning models that are almost 20 times larger, as stated by the Institute of Foundation Models at MBZUAI.
The new system is ranked among the world’s best reasoning models and excels in various math benchmarks, surpassing all open-source models in AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD.
Features of K2 Think
K2 Think is built upon six foundational pillars of innovation, including chain-of-thought supervised fine-tuning, reinforcement learning with verifiable rewards, agentic planning, and scalability during testing.
This model will also be accessible on Cerebras’ wafer-scale computing platform, optimized for speculative decoding at a speed of 2,000 tokens per second.
Khaldoon Khalifa Al Mubarak, the chairman of MBZUAI’s Board of Trustees and a member of the Artificial Intelligence and Advanced Technology Council (AIATC), remarked, “The standard set by K2 Think creates a new global benchmark and highlights the innovative leadership of MBZUAI’s Institute of Foundation Models. This initiative exemplifies the UAE’s dedication to developing advanced systems through its institutions and making them available globally.”
Peng Xiao, a board member of MBZUAI and the group CEO of G42, added, “K2 Think shifts the AI reasoning paradigm from ‘bigger is better’ to ‘smarter is better’. By demonstrating that smaller, resource-efficient models can compete with the largest reasoning architectures, we are embarking on a new wave of AI innovation.”
In a notable departure from typical practices, K2 Think is fully open source, encompassing not only model weights but also training data and deployment software.
Professor Eric Xing, the president of MBZUAI, described the system as ushering in “a new age of affordable, reproducible, and accountable AI.”
MBZUAI’s Open-Source Model Collection
K2 Think expands MBZUAI’s open-source model collection, which already includes Jais (Arabic), NANDA (Hindi), and SHERKALA (Kazakh), and builds upon K2-65B, which was the first fully reproducible open-source foundation model released in 2024.
K2 Think can be accessed at K2Think.Ai/K2Think as well as on Hugging Face.