MBZUAI’s Institute of Foundation Models has launched six K2 Horizon AI models, ranging from 0.9 billion to 375 billion parameters, with the weights, code, training data and methodology released openly.
The Abu Dhabi-based lab says the K2 Horizon AI models are available immediately through Hugging Face, vLLM and SGLang. Their code and model weights are released under the Apache 2.0 licence, while APIs are available through inference partners including Compass, Cerebras and Nebius.
K2 Horizon AI models cover watches, phones and enterprise servers
The fleet starts with a 0.9B model intended for constrained devices such as watches and glasses. The 3.7B and 7B models target phones and other on-device applications, while the 32B dense model is built for local hosting on laptops and on-premise servers.
The sparse 36B-A4B model activates 4B parameters and uses the lab’s Mixture of Value Attention architecture. At the top is the 375B-A23B model, with 23B active parameters, aimed at larger reasoning and agentic workloads. The vLLM deployment recipes list the flagship with 379.17B stored parameters and a 131,072-token context window.
IFM says all six models share core architecture, training methods, interfaces and deployment tooling. Its dynamic model-routing system can send a task to the lower-cost model best suited to it, allowing developers to move between sizes without changing their workflow.
Why the release is more open than the usual model drop
IFM says it is releasing training code, data-construction recipes or training data, checkpoints and evaluation information alongside the final models. That gives researchers more to inspect than a downloadable set of weights, although reproducing a frontier model still depends on access to substantial computing resources.
“Open source is much more than open weights. Science works when others can see the data, follow the method, reproduce the result, and improve on it.”
Eric Xing, founder of IFM and president of MBZUAI
The release follows MBZUAI’s K2 Think hackathon and its work helping students launch AI start-ups. IFM operates from Abu Dhabi, Silicon Valley and Paris, while the models themselves are available to developers worldwide. K2 Horizon broadens the university’s open-model work across six sizes.
Who created K2 Horizon?
K2 Horizon was created by the Institute of Foundation Models, a global AI research lab launched by Mohamed bin Zayed University of Artificial Intelligence in May 2025. IFM has offices in Abu Dhabi, Silicon Valley and Paris.
What does 36B-A4B mean?
The 36B-A4B model has about 36 billion stored parameters but activates about 4 billion for each token. It uses a mixture-of-experts design to reduce the computation needed for each response.
What does the Apache 2.0 licence allow?
Apache 2.0 is a permissive open-source licence that generally allows people to use, modify and distribute the software, subject to its licence conditions and notices.
