The six-model release gives researchers access to weights, code, training data and methodology, offering a more complete look at how frontier AI systems are built. 

[For more news, click here]

The debate around open AI has become less straightforward as more companies release models that can be downloaded and run outside their own platforms. Model weights may be public, but the data and methods used to train those models often are not. The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) is taking a broader approach with K2 Horizon, a new family of six models that puts much more of the underlying work in public view. 

Released on September 3, K2 Horizon spans 0.9 billion to 375 billion parameters, with models intended for everything from highly constrained devices to enterprise systems. Alongside the weights, IFM has released source code, training data and methodology under the Apache 2.0 licence. Researchers can therefore study how the models were developed, while developers can adapt and deploy them without relying solely on an API. IFM describes K2 Horizon as the largest fully open model release in AI history. 

The distinction between open weights and a more complete release is becoming increasingly relevant to researchers. A downloadable model can be useful without revealing much about the process that produced it. Training data and methodology provide another layer for researchers to examine, although reproducing a frontier model still requires substantial computing resources. K2 Horizon follows the direction IFM has taken with its earlier K2 models but expands that approach across a much wider range of model sizes. The size range also reflects the different ways companies are beginning to use AI. The 0.9B model is designed for devices such as watches and glasses, while the 3.7B and 7B models are aimed at phones and other local applications. The 32B model targets local hosting and on-premise servers, and the 36B-A4B version activates four billion parameters despite having 36 billion parameters in total. At the top sits the 375B-A23B model, with 23 billion active parameters, intended for more demanding reasoning and agentic workloads. IFM says the three smallest models achieved state-of-the-art results in their respective size categories across mathematics, reasoning, software engineering and agentic tasks. Those claims will now have to face independent testing and real-world use. 

Several of the technical decisions behind K2 Horizon are aimed at reducing the cost of running capable models. IFM says its diffusion distillation technique generates blocks of tokens in parallel and can improve model performance by roughly three times without degrading response quality. Its Mixture of Value Attention architecture is intended to improve reasoning without adding more computation, while dynamic model routing can direct different tasks to models of different sizes. The practical value of those methods will depend on how they behave outside controlled evaluations, particularly when businesses have to balance performance with hardware, latency and operating costs. 

“We believe meaningful AI progress depends on the ability to examine, build upon, and improve the technology, not simply access it through an API. The most important technology of our time should be built with the world, not kept from it. K2 Horizon is what that conviction looks like when an institution commits to it completely, and it reflects the environment we work in here in the UAE: a country that decided to build AI rather than wait for it.” said Professor Eric Xing, Founder of IFM, and President and University Professor of MBZUAI 

IFM was established by MBZUAI in May 2025 and now operates across Abu Dhabi, Silicon Valley and Paris. Its expansion comes as the UAE continues to build domestic research and technical capabilities around AI. The university itself has grown from a graduate research university into an institution with undergraduate education as well, while the UAE government has also introduced programmes aimed at developing advanced science and technology talent. 

K2 Horizon gives that effort a model family that outside researchers can actually examine and use. The release will ultimately be judged by what happens after the launch. Researchers can test the training claims; developers can see how the models perform in their own applications and businesses can decide whether the systems are competitive enough to justify adoption. For IFM, putting the work in public makes those judgments possible without asking users to take the institute's word for it. In an AI industry where openness is often defined by how much of a finished model can be downloaded, giving researchers access to the work behind it is a stronger standard to set. 

Related Articles 

DeepSeek Nears $74 Billion Valuation Ahead of Potential 2027 IPO 

GenBio Unveils AIDO Cell, an AI ‘World Model’ of the Human Cell 

UAE Launches National Program for Advanced Sciences and Technologies to Build Homegrown Talent