The UAE has launched K2 Think, a powerful open-source AI model that aims to rival giants like OpenAI and China’s DeepSeek by offering advanced reasoning capabilities in a compact and efficient package.
Quick Summary – TLDR:
- K2 Think is a 32-billion-parameter AI model from the UAE, optimized for math, code, and science
- It outperforms much larger models like DeepSeek R1 and matches proprietary systems like ChatGPT in key benchmarks
- The model is designed for speed and accessibility, delivering 2,000 tokens per second using Cerebras hardware
- It is fully open source, allowing developers and enterprises to use, modify, and deploy it freely under Apache 2.0 license
What Happened?
A research collaboration between the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) and Emirati tech firm G42 has resulted in the launch of K2 Think, a cutting-edge reasoning model built to rival leading systems from the United States and China. Announced in Abu Dhabi, the model is being positioned as the fastest open-source AI system available and represents a significant strategic step in the UAE’s Artificial Intelligence 2031 plan.
Introducing K2 Think – a breakthrough in advanced AI reasoning.
— MBZUAI (@mbzuai) September 9, 2025
Developed by MBZUAI’s Institute of Foundation Models and @G42ai, K2 Think delivers frontier reasoning performance at a fraction of the size of today’s largest systems.
Smaller. Smarter. Open to the world.… pic.twitter.com/l3MlOcH30o
Small but Smart: The UAE’s Efficient AI Breakthrough
Unlike traditional AI models that scale up with massive parameter counts, K2 Think takes a smarter path. It operates with just 32 billion parameters, a fraction of what powers models like DeepSeek R1’s 671 billion. Yet, it matches or even beats them in benchmark tests across math, code, and science.
The performance is backed by real data. K2 Think scores:
- 90.8 on AIME 2024
- 81.2 on AIME 2025
- 73.8 on HMMT 2025
- 60.7 on OMNI-MATH-HARD
- 64.0 on LiveCodeBench v5 (code)
- 71.1 on GPQA-Diamond (science)
Blazing Fast: 2,000 Tokens per Second
One standout feature of K2 Think is its speed. The model can generate 2,000 tokens per second, more than 10 times faster than most current AI models like Google’s Gemini Flash, which sits at around 258 tokens per second. This ultra-fast performance is powered by Cerebras’ Wafer-Scale Engine, which processes long 32,000-token responses in just 16 seconds for a task that takes over two minutes on traditional GPUs.
Built for Reasoning, Not Just Chat
K2 Think is engineered not just for conversation but for advanced reasoning. It uses six core methods to optimize performance:
- Supervised Fine-Tuning with complex chain-of-thought examples
- Reinforcement Learning with Verifiable Rewards
- Agentic Planning for structured reasoning
- Test-Time Scaling using Best-of-N sampling
- Speculative Decoding to speed up inference
- Hardware Optimization via Cerebras technology
These approaches allow the model to plan and execute step-by-step solutions, especially in technical fields like mathematics and scientific research.
Fully Open Source and Commercial-Ready
K2 Think is available under an Apache 2.0 license, meaning developers and enterprises can download, adapt, and deploy the model without restrictions or fees. Everything from model weights and training data to inference tools and fine-tuning code is accessible at k2think.ai and Hugging Face.
The project also scored well in internal safety evaluations, including:
- 0.83 in high-risk content refusal
- 0.89 in conversational robustness
- 0.72 in jailbreak resistance
- 0.56 in cybersecurity and data protection
A Strategic Shift in Global AI Dynamics
This launch is part of a broader push by the UAE to establish itself as a serious player in global AI. The government-backed Artificial Intelligence 2031 initiative seeks to boost local innovation while also attracting global partnerships. The UAE is already a founding partner in Trump’s Project Stargate and is investing heavily in data infrastructure with companies like Microsoft, Nvidia, Cisco, and Oracle.
K2 Think also highlights the UAE’s ambition to diversify beyond oil, using AI as a new pillar of economic growth and influence. “We can use limited resources to make things work,” said MBZUAI President Eric Xing, underscoring the country’s focus on cost-effective innovation.
SQ Magazine Takeaway
I love how K2 Think flips the script on the old belief that only giant models can lead in AI. What the UAE has done here is show the world that brains can beat brawn. They took a compact system and tuned it so well that it’s beating some of the most resource-heavy models out there. The open-source part is the cherry on top. This model isn’t just a tech feat, it’s a strategic power move. And if this is what they’re doing in 2025, the rest of the world better keep up.
