In Abu Dhabi, it is home to Mohamed Bin Zayed University of Artificial Intelligence. This is an example of an increasing investment in the UAE to establish leadership positions in AI.
Getty
Mohamed Bin Zayed University of Artificial Intelligence and G42 Technology Group are both headquartered in Abu Dhabi within the United Arab Emirates, but have announced that the entirely open source K2 thinks of AI inference models. It is an interesting technical achievement in itself due to its potential utility across different domains, as well as how the model was developed. More than that, it fits the important themes I’ve been keeping an eye on this year. When it comes to AI, it is a serious commitment of multiple governments in the Arabian Peninsula.
Over the past few months, leaders of AMD, Nvidia and Qualcomm have regained major development contracts after accompanying President Trump on his local tour. Saudi Arabia has demonstrated its own severity through its state-backed Humain AI development company. And my own research touches on this across several different topics, including my recent assessment of US export controls regarding Nvidia.
Currently, K2 offers new wrinkles on how to build highly efficient AI inference models, suggesting progress that relies on thorough beliefs about open source technologies.
How to punch a 32 billion parameter inference model above its weight
When it comes to AI innovation, the Mohamed Bin Zayed University of Artificial Intelligence may benefit from its novelty as Emirati’s wealth supports it. Certified in 2020, Mbzuai currently offers one bachelor’s degree (BS in AI) along with numerous master’s and doctoral programs in computer science, machine learning, and robotics. And it seems they are committed to a fresh approach to AI based on what I saw.
We consider the K2 to be a small but powerful and innovative reasoning system, sitting on a base model of similar size (QWEN2.5-32B). With 32 billion parameters, K2 performs faster in mathematical inference than models that have several orders of magnitude larger, such as the Deepseek R1.
Developers at Mbzuai’s Foundation Model Institute are promoting six “pillars of innovation” that allow for high performance even in this small model.
Long-minded monitoring facility reinforcement learning with verifiable reward agent plans before scaling hardware that makes inference based on test time scaling inference.
Those who pique their appetites on technical details can learn more in research papers that IFM has been published in conjunction with the release of the model. I would like to focus on just a few of the particular pillars before I pull back to see the larger picture.
First, the agent plan before inference – as the label suggests – applies agent AI to reconstruct the important concepts from the input prompt before inference itself is made. As far as the researchers could convey, this is their novel contribution, and it makes a lot of sense to me. There is evidence from cognitive science that our human brains engage in a small number of different plans than our main reasoning departments, but they can help us to think about the issue more clearly. Smart.
As a former longtime semiconductor executive, I would also like to draw attention to the inference-optimized hardware pillars. K2 uses Celebras wafer-scale engine system, which the company advertises as “the world’s largest and most powerful AI chip.” Celebras will build these WSE systems across the wafer, providing 19 times more transistors (4 trillion transistors (4″ and 28 times more petaflops than the NVIDIA B200 GPU. If you are trying to expand this, how quickly can you get the number of Celebras WSE-3s? I was mainly impressed by the experts at Mbzuai who are using such a high horsepower silicon solution.
In the big picture, I think what we’re looking at here is an even version of what we saw from Deepseek’s first Bombshell announcement in early 2025. This uses an innovative combination of techniques above the total of the parts, resulting in much better performance than expected.
K2’s benchmark performance considers inference models and other important AI models
Mohamed Bin Zayed University of Artificial Intelligence, Foundation Model Research Institute
And how good is the performance? Internal testing shows that K2 is considering performance comparable to some of the best open source models in benchmarks such as AIME24/25, GPQA-Diamond, and HMMT. However, all peers on these benchmarks require several times more parameters. And models that truly surpass k2 (e.g. GPT-5 and GEMINI 2.5) are on the order of 20x. Of course, I think K2 is public, so I would like to study independent external benchmarks, but from what I’ve seen so far, the model approach and efficiency is very appealing.
Open source for inference models.
The K2 has the predecessor model K2-65B, released in 2024. The developers call the model “the world’s first fully reproducible open source foundation model,” and take it a step further here. The K2-65B recorded the prerequisites for the 6.5 billion parameter model from end to end. Currently, K2 documents post-inference paths for inference, with the goal of not only reproducing what a model can do, but also allowing anyone to replicate how it learns to infer. This is promising to be within the “black box” of generator AI.
One of the internal documents read during the briefing stage is that K2 is “the first open source inference system that reveals the importance of each step.” So, if you choose to use this model, you can name it by, for example, getting it by hugging your face. It’s a complete deployment system. The details of the data are big deals. This is because many “open source” models do not disclose this.
This is especially interesting to me. Because I have been observing open and unique technology models for 30 years. Not in either case, but in most cases openness is a win over the long term. It arises from the perspective of widespread adoption, ecological robustness, and full prosperity, and many other metrics are certain. So I’m eager to look at the level of commitment to open source by the various organizations that are building and running AI models. For now, it’s enough to praise K2’s philosophical approach to openness that animates K2’s ideas.
Arabian countries’ ambitions in AI are worth seeing
We consider K2 to be the latest in a range of open source AI models developed in the UAE. In addition to the K2-65B, there are Jais (promoted as the most advanced LLM in the Arabic world), Nanda (Hindi) and Sherkala (Kazakh). And the developers say this year’s high-performance model release from the Foundation Model Research Institute. Incidentally, the IFM itself, which was only released in May 2025 and obviously ran to the ground, is structured to support multiple sites, in this case Abu Dhabi, Paris and Silicon Valley distributed teams. The whole thing is built to move faster.
One internal explanation for K2 is called “as much as the model” because of its end-to-end openness and reproducibility. I also think its rapid development and launch shows that there is a deeper business model, or industrial leadership model, in play for the UAE. Like its Qatar and Saudi Arabia neighbours, the Emirati government appears to have died in asserting leadership positions in the open source AI model.
These Arabian countries have some serious advantages that should not be discounted. It’s an enormous amount of funding resources and effective, unlimited access to inexpensive energy for the extremely important AI data center. I was fascinated by the work that emerged from these centers. And they bet that the sky is at its limit, what other organizations in the region may achieve in the coming years.
Moor Insights & Strategy provides or provides paid services to technology companies, including all tech industry research and analyst companies. These services include research, analysis, advice, consulting, benchmarking, acquisition matchmaking, video, and speaking sponsorship. Of the companies mentioned in this article, Moor Insights & Strategy currently has (or has) a paid business relationship with AMD, Nvidia, and Qualcomm.
