UAE Unveils Lightweight, Open-Source AI Model K2 Think

The United Arab Emirates introduced K2 Think on September 9, 2025, marking the country's boldest move yet on the global artificial intelligence stage. This open-source reasoning model, developed by the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and technology group G42, achieves performance levels comparable to much larger AI systems using just 32 billion parameters, according to Emirates 24/7.

K2 Think represents what researchers call a paradigm shift from the "bigger is better" to the "smarter is better" approach in AI development. Eric Xing, president of MBZUAI, stated that the model "outperforms flagship reasoning models that are 20 times its size," describing it in a WIRED interview as "a technical breakthrough, and in my view, a game-changing shift."

Technical Innovations Drive Efficiency

The system's breakthrough relies on six technical pillars that enable outstanding performance despite its compact size. These include training with long-chain reasoning for deeper logical inference, reinforcement learning with verifiable rewards to enhance accuracy, and agentic planning that allows the model to break complex tasks into stages for logical analysis. Scaling techniques during testing further boost adaptability, while speculative decoding, optimized for Cerebras hardware, delivers unprecedented throughput.

Running on Cerebras wafer-scale processors, K2 Think achieves an impressive speed of 2000 tokens per second, making it one of the fastest and most efficient reasoning systems globally. According to WIRED, this combination enables the model to rival or surpass the performance of OpenAI and DeepSeek models with over 200 billion parameters, while requiring significantly fewer computational resources. The model leads among all open-source systems in mathematical reasoning benchmarks, including AIME '24/'25, HMMT '25, and OMNI-Math-HARD.