What MBZUAI released
K2 Horizon is presented by Mohamed bin Zayed University of Artificial Intelligence as a fleet of six AI foundation models ranging from 0.9 billion to 375 billion parameters. MBZUAI says every model is released fully open, including weights, code, training data and methodology. WAM reports the same core details and says the release enables researchers and developers to inspect, reproduce and adapt the models.
The release is not described only as a future plan or private pilot. Both approved sources say K2 Horizon is available immediately through Hugging Face, vLLM and SGLang. They also say the K2 Horizon application programming interface is available through inference partners including Compass, Cerebras, AWS and Nebius. MBZUAI says the models and code are released under the Apache 2.0 license.
For business readers, the practical point is that K2 Horizon is a released model family rather than a single model announcement. The smallest model has 0.9 billion parameters. The largest has 375 billion parameters, with 23 billion active parameters. Between those ends, the family includes models sized at 3.7 billion, 7 billion, 32 billion and 36 billion parameters, with the 36 billion model activating 4 billion parameters. That range gives technical teams several sizes to test against different constraints.
Why the range matters
MBZUAI says the 0.9 billion parameter model is designed for highly constrained environments such as watches and glasses. It also says the 3.7 billion and 7 billion parameter models bring advanced capabilities to phones and other applications that run on the device itself. WAM reports the same placement of smaller models for constrained and device based settings.
The larger models are positioned differently. MBZUAI describes the 32 billion model and the 36 billion model with 4 billion active parameters as options for local hosting and servers located on premises. The 375 billion model with 23 billion active parameters is described as the flagship model for demanding enterprise deployments. WAM says the flagship is designed for demanding enterprise workloads and for competition with leading open weight models in reasoning and agentic tasks.
UAE organisations should treat those statements as the vendor and research institution position, not as proof that every model is production ready for every sector. The sources support availability and intended deployment categories. They do not replace internal testing for accuracy, latency, security, cost, integration needs, governance or sector regulation. A careful evaluation can still use the published model sizes and access terms as a starting point.
What openness changes
The approved sources place openness at the centre of the release. MBZUAI says K2 Horizon goes beyond releases that publish only model weights by also providing code, training data and methodology. This matters for organisations that need more than an external application programming interface before they can make decisions about AI use in sensitive or regulated workflows.
For UAE technology leaders, the release creates a clearer evaluation path. Teams can compare small models for edge settings, medium models for local hosting and the largest model for enterprise reasoning workloads. Because MBZUAI says the six models largely share architecture, vocabulary, training methodology, interfaces and deployment tooling, technical teams may be able to compare model sizes with less variation in surrounding tooling. The MBZUAI source notes that the 0.9 billion model uses a smaller vocabulary.
MBZUAI also describes technical work inside the family. It says diffusion distillation can improve model speed by roughly three times without reducing response quality. It describes Mixture of Value Attention as an architecture intended to improve reasoning without additional computation. It also describes dynamic model routing as a technique that directs tasks to the most cost effective model. These are MBZUAI claims and should be tested in each organisation’s own workload before procurement or deployment decisions.
Relevance for UAE organisations
K2 Horizon has specific local significance because IFM operates from Abu Dhabi and is part of MBZUAI. WAM reports that IFM was established in May 2025 and operates from Abu Dhabi, Silicon Valley and Paris. MBZUAI also says the institute employs AI researchers, scientists and developers across those offices.
The release may be especially relevant to UAE organisations that are deciding where AI should run. Some use cases may favour small models close to the device, especially where connectivity, response time or local processing matter. Other use cases may fit local servers or on premises infrastructure. More demanding reasoning workloads may need evaluation against the largest model or against other available systems.
The approved sources do not state that UAE organisations have already deployed K2 Horizon in production. They also do not state that K2 Horizon is the right choice for every business process. What they do support is narrower and useful. K2 Horizon is available now, it is released with open components that MBZUAI identifies, and it gives organisations a model family to examine across edge, local and enterprise settings. That makes it a practical item for UAE artificial intelligence planning, provided each team treats the announcement as the beginning of technical validation rather than the end of it.
Plan Your AI Project with Confidence
Discuss your goals, current systems, and practical opportunities with ElephantClock Technology. We will help you identify a focused and responsible path for your AI project.
Get a Free Consultation