Introduction to K2 Think
On September 9, 2025, the Institute of Foundation Models at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42 unveiled
K2 Think, an exceptional open-source system that revolutionizes AI reasoning. This advanced model is not only designed for optimal performance but also emphasizes parameter efficiency, making it a significant contender in AI technology.
Performance and Efficiency of K2 Think
K2 Think stands out in its category by delivering performance comparable to the most advanced reasoning models, yet at a fraction of the size. With
only 32 billion parameters, it surpasses flagship models that are 20 times larger. This impressive leap in parameter efficiency redefines what's possible for compact architectures, setting a new benchmark for AI development.
Innovative Foundations of K2 Think
This innovative system is built upon six core pillars, leading to a new class of reasoning models. K2 Think employs a long sequence of thoughts for supervised fine-tuning, enhancing its logical depth. Following this, reinforcement learning based on verified rewards boosts accuracy when tackling complex problems. The adoption of agent-based planning allows the model to deconstruct intricate challenges before processing them effectively. Additionally, scalability techniques implemented during testing further enhance adaptability, facilitating seamless real-world applications.
Unmatched Throughput and Speed
One of the most striking features of K2 Think is its integration with Cerebras' wafer-scale, inference-optimized computing platform. This collaboration means K2 Think will allow researchers and innovators to push performance boundaries at unprecedented speeds. Harnessing optimized speculative decoding for Cerebras hardware, K2 Think achieves an incredible throughput of
2000 tokens per second, solidifying its status as one of the fastest and most efficient reasoning systems available today.
Leading the Industry in Performance
K2 Think doesn't just excel in operational efficiency; it leads the industry in performance metrics for various recognized competitions including AIME '24/'25, HMMT '25, and OMNI-Math-HARD. By establishing itself as a frontrunner among open-source models, K2 Think exemplifies how open innovation can fuel significant advancements in AI technology.
Positioning AI in the UAE
This innovative leap forward is more than a technical achievement; it marks a pivotal chapter in AI development within the
United Arab Emirates. It showcases how open innovation, coupled with robust public-private partnerships, can position Abu Dhabi as a global hub for artificial intelligence. The impact is clear: the future of AI reasoning is not solely defined by size, but also by creativity and collaboration.
His Excellency Khaldoon Khalifa Al Mubarak, Chairman of the MBZUAI Board of Trustees, highlighted the significance of K2 Think by stating that it underscores the groundbreaking excellence of the MBZUAI Foundation Models initiative. It paves an accelerated path for global collaboration and top-tier research.
A Paradigm Shift in AI Development
Peng Xiao, a board member of MBZUAI and Group CEO of G42, articulated that K2 Think represents a paradigm shift from "bigger is better" to "smarter is better." Supported by the UAE's thriving tech ecosystem, this initiative challenges the notion that only large models can perform at the highest levels. It marks the start of a new wave of AI innovation that is leaner and more resource-efficient.
A Commitment to Transparency
Unlike most open AI models that limit access to just model weights, K2 Think embraces full open-source principles by making everything from training data to software code publicly available. This unprecedented level of transparency ensures that every learning process can be scrutinized, replicated, and expanded upon by global research communities, fostering a rich environment for collaboration and innovation in AI.
The Future of AI Research
Professor Eric Xing, President of MBZUAI, remarked that K2 Think signifies a major leap for the global AI research community. By embodying efficiency in a fully transparent framework, it heralds a new era of accountable, reproducible AI. Even at just five years old, MBZUAI showcases its commitment to advancing global scientific and technological paradigms through innovative efforts by its dedicated researchers and engineers.
Expanding the Open-Source Landscape
K2 Think is a continuation of a growing portfolio of open-source models developed in the UAE, including Jais, NANDA, and SHERKALA. This follows the groundbreaking K2-65B release, which was the world’s first fully reproducible open-source foundation model, launched in 2024.
For those eager to explore what K2 Think has to offer, it is now available at
K2 Think and on
Hugging Face.
About MBZUAI
The
Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) is a research-oriented institution located in Abu Dhabi, dedicated to advancing scientific knowledge in the field of AI. By providing top-tier training and interdisciplinary research, MBZUAI seeks to empower the next generation of AI leaders and drive impactful innovations. In 2025, the university launched its inaugural undergraduate program, a Bachelor of Science in AI, encompassing both business and technology tracks. To learn more about MBZUAI, visit
mbzuai.ac.ae.