The Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and G42 AI announced the launch of K2 Think, an open-source system for advanced AI reasoning.

K2 Think introduces a new approach to building smarter and more efficient AI. With just 32 billion parameters, it outperforms flagship reasoning models that are 20 times larger. This breakthrough in parameter efficiency makes it a powerful alternative for advanced reasoning.

Built on six pillars of innovation, K2 Think employs long chain-of-thought supervised fine-tuning to improve logical depth. It also applies reinforcement learning with verifiable rewards to sharpen accuracy. Agentic planning allows the model to break down complex challenges, while test-time scaling boosts adaptability.

The system will be available on Cerebras’ wafer-scale platform, optimized for inference. With speculative decoding tailored for Cerebras hardware, K2 Think is expected to reach 2000 tokens per second. This positions it among the fastest and most efficient reasoning systems.

K2 Think ranks among the industry’s top reasoning systems and leads all open-source models in math performance across AIME ’24/’25, HMMT ’25, and OMNI-Math-HARD.

• Outperforms models 20x larger with fewer parameters
• Achieves 2000 tokens per second on Cerebras hardware
• Leads open-source benchmarks in reasoning performance

Leaders emphasized its global and regional significance. His Excellency Khaldoon Khalifa Al Mubarak, Chairman of MBZUAI’s Board of Trustees, said K2 Think sets a new global benchmark. He highlighted the UAE’s role in developing scalable innovations with worldwide impact.

Peng Xiao, MBZUAI Board Member and Group CEO of G42, reported that K2 Think shifts the AI reasoning paradigm from “bigger is better” to “smarter is better.” He added that UAE’s ecosystem is advancing AI with technology that is efficient, open, and highly capable.

Unlike most open AI models that release only weights, K2 Think is fully open source, including training data, parameter weights, and deployment software. This transparency enables the global research community to study, reproduce, and extend its methods.

Professor Eric Xing, President of MBZUAI, revealed that K2 Think represents a significant advancement for global AI research. He said it promotes cost-effective, reproducible, and accountable AI within a transparent framework.

K2 Think also extends the UAE’s open-source AI portfolio, which includes Jais, NANDA, SHERKALA, and K2-65B. Together, they highlight Abu Dhabi’s ambition to lead in artificial intelligence through ingenuity and collaboration.

Related post

View all