Artificial Intelligence – UAE & MBZUAI

UAE makes a play for open source.

• The UAE has launched a new model aimed at providing an alternative to China in open source, but the quality of this model is not immediately clear.
• Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) has released K2 Think, a 70bn parameter “reasoning” model that it has made available to open source in its entirety.
• MBZUAI is a state-financed university aimed at developing AI and, in some respects, is similar to Tsinghua University in Beijing, which has been involved in some of China’s most important technology innovations.
• Being open source in this instance means not just the model and its weights but also the data that was used, as well as the training history to ensure that the model is as open as possible.
• MBZAUI makes some bold claims for K2 Think with leadership on select benchmarks, as well as the claim that the underlying LLM (K2 V2) outperforms Alibaba’s Qwen2.5 72B and approaches the performance of the much larger Qwen3-235B.
• The model has been launched in conjunction with G42 and Cerebras Systems (see here), meaning that K2 Think will be available on Core 42’s cloud with inference being carried out on Cerebras silicon.
• Crucially, it will also be freely available for download so that anyone who wants to use it can run it on their own terms and within their own infrastructure.
• MBZUAI shows benchmarks against older versions of K2 (see here) where it sees steady improvement, and also employed Artificial Analysis to run benchmarks against the other open source models in its class.
• As one would expect, it fares pretty well against its peers, but the independence of the study will be questioned because the study was paid for by MBZUAI.
• On my own simple tests, I get mixed results, but the model is blazingly fast.
• When asked to generate a table of Nvidia data centre chip generations and their key specifications, the model fails badly, and the result is useless.
• However, when asked to explain the steps regarding how peanut butter is made in a factory, it does a passable job, although it failed to draw a diagram of the colloidal mill that is used to make peanut butter very smooth.
• Although the results are mixed, the model is blazingly fast, many times faster than the already fairly fast ChatGPT Pro and many orders of magnitude faster than the glacial Gemini 3 Pro.
• This could be down to several factors such as the small model size (70bn vs trillions), Cerebras silicon, which is optimised for fast inference and server load, as this model is almost unknown, and so the server load may be low.
• I think that the real purpose of this model is to offer an alternative to China, which has taken over leadership in the open source after Meta appeared to drop out.
• So strong is China’s dominance that US start-ups that are using open source models as their starting point are almost all using Chinese models.
• In the worsening geopolitical climate, a non-affiliated or neutral alternative should attract a lot of interest as MBZUAI has no real skin in the game other than to make the best model possible.
• If it can gain some traction with developers and start-ups, then the UAE’s reputation as a centre of competence in AI will begin to take off.
• This, in turn, will be extremely beneficial when it comes to filling the 5GW of compute capacity that the UAE intends to build over the next several years in its AI campus.
• I suspect that the UAE (like Saudi Arabia) will use its abundant real estate, permissive regulatory environment and cheap energy to compete on price for customers looking for compute, but if this model lives up to its billing, then it may be able to also compete on quality.
• I remain pretty positive on the outlook for this new industry in the Middle East and expect that it will become a significant provider of compute, in the same way that it is a significant provider of energy today.
• If MBZUAI can put some icing on the cake, then the UAE’s aspirations will have an even better chance of being realised.