K2 Think sets new AI reasoning benchmark with compact model
The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42 have launched K2 Think, an open-source system for advanced AI reasoning, designed to offer high performance with a compact architecture.
K2 Think has been developed using 32 billion parameters and outperforms established flagship models that operate with up to 20 times more parameters. The system's parameter efficiency marks a notable advancement for reasoning tasks, allowing for significant computational savings compared to larger AI systems.
According to MBZUAI and G42, K2 Think is underpinned by six specific areas of technical focus: chain-of-thought supervised fine-tuning for greater logical complexity, reinforcement learning with verifiable rewards to improve accuracy, agentic planning for decomposing and addressing complex problems, and test-time scaling for adaptability. The model will also soon be accessible through Cerebras' wafer-scale hardware, promising researchers and AI developers substantially increased throughput rates of up to 2000 tokens per second.
K2 Think has achieved strong results in industry-standard mathematical reasoning benchmarks, outperforming other open-source models in evaluations such as AIME '24/'25, HMMT '25, and OMNI-Math-HARD.
Industry perspective
"The new global benchmark set by K2 Think underscores the pioneering excellence of MBZUAI's Institute of Foundation Models initiative, an expedited pathway for global collaboration and cutting-edge research. It is also an example of the UAE's commitment to building advanced systems that are developed by our institutions and shared with the world - ultimately progressing technically groundbreaking, practical, and scalable innovations with transformative global impact."
This statement was made by His Excellency Khaldoon Khalifa Al Mubarak, Chairman of MBZUAI's Board of Trustees and Member of the Artificial Intelligence and Advanced Technology Council (AIATC), highlighting the significance of the launch for both MBZUAI and the wider UAE.
K2 Think also reflects a shift in the approach to AI system development. Peng Xiao, MBZUAI Board Member, Council Member of Abu Dhabi's AI and Advanced Technology Council, and Group CEO of G42, stated:
"K2 Think has shifted the AI reasoning paradigm from 'bigger is better' to 'smarter is better'. MBZUAI, supported by the UAE ecosystem, is pushing the AI frontier with technology that is open, efficient and highly capable. By proving that smaller, more resourceful models can rival the largest reasoning systems, this milestone marks the beginning of the next wave of AI innovation."
Openness and transparency
K2 Think differentiates itself from many open-source models by providing openness not only with model weights but also with full training data, deployment code, and test-time optimisation tools. The developers say this level of transparency will support reproducibility and further research by providing comprehensive resources to the global AI community.
Professor Eric Xing, MBZUAI President and University Professor, commented on the importance of this open approach:
"K2 Think, developed by MBZUAI's Institute of Foundation Models, is a significant advancement for the global AI research and development community. By delivering these advances in a fully transparent framework, we are ushering in a new era of cost-effective, reproducible and accountable AI. For an institution just five years young, we are immensely proud of our global researchers, engineers, and teams who are advancing science and technology with ingenuity and a pioneering spirit."
Context and legacy
K2 Think joins a set of AI models developed in the UAE, including Jais for Arabic, NANDA for Hindi, and SHERKALA for Kazakh languages, as well as K2-65B, described as the first fully reproducible open-source foundation model launched in 2024. The organisations expect that K2 Think's compact but capable architecture and transparent release will support further research and the practical deployment of high-performance reasoning models worldwide.