MBZUAI and G42 Launch K2 Think: A Breakthrough Open-Source AI Reasoning System

The Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and technology group G42 have unveiled K2 Think, a groundbreaking open-source system for advanced AI reasoning.

K2 Think represents a major leap in AI efficiency and intelligence. Despite having just 32 billion parameters, it outperforms flagship reasoning models that are 20 times larger, setting a new standard for compact yet powerful AI architectures.

The system is built on six pillars of innovation. It employs long chain-of-thought supervised fine-tuning to enhance logical reasoning, reinforcement learning with verifiable rewards to improve accuracy on complex problems, and agentic planning to break down intricate challenges before tackling them. Additional test-time scaling techniques further boost adaptability. Soon, K2 Think will be accessible on Cerebras’ wafer-scale, inference-optimized computing platform, delivering speeds of up to 2,000 tokens per second, making it one of the fastest and most efficient reasoning systems globally.

K2 Think leads all open-source AI models in mathematics performance across AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD, demonstrating its advanced capabilities.

Beyond its technical achievements, K2 Think reflects the UAE’s growing prominence in AI innovation. It showcases the potential of open collaboration between public institutions and private companies, positioning Abu Dhabi as a global hub for AI research.

His Excellency Khaldoon Khalifa Al Mubarak, Chairman of MBZUAI’s Board of Trustees, said:
"K2 Think sets a new global benchmark, highlighting MBZUAI’s pioneering excellence. This initiative demonstrates the UAE’s commitment to developing advanced systems that are shared with the world, fostering innovations with transformative global impact."

Peng Xiao, MBZUAI Board Member and CEO of G42, added:
"K2 Think shifts the AI reasoning paradigm from ‘bigger is better’ to ‘smarter is better’. By showing that smaller, resource-efficient models can rival the largest systems, we are entering a new wave of AI innovation."

Unlike most open AI models, K2 Think is fully open source, including training data, parameter weights, deployment software, and test-time optimization tools. This transparency allows researchers worldwide to study, reproduce, and expand the model’s capabilities.

Professor Eric Xing, MBZUAI President, said:
"K2 Think represents a significant advancement for global AI research. By providing these innovations in a fully transparent framework, we are promoting cost-effective, reproducible, and accountable AI."

K2 Think builds on a series of UAE-developed open-source models, including Jais (Arabic), NANDA (Hindi), SHERKALA (Kazakh), and follows the legacy of K2-65B, the world’s first fully reproducible open-source foundation model released in 2024.