MBZUAI launches K2 Horizon family of six open AI models
New models range from 0.9 billion to 375 billion parameters, targeting applications from wearable devices to enterprise AI
03 September, 2026
TT
16
The Institute of Foundation Models (IFM) at Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) has launched K2 Horizon, a family of six artificial intelligence foundation models ranging from 0.9 billion to 375 billion parameters.
MBZUAI said all six models are being released with their weights, code, training data and methodology available, allowing researchers, developers and institutions to inspect, reproduce and adapt them.
The university described K2 Horizon as the largest fully open model release in AI history, based on the breadth of models and supporting resources being made available.
The family includes models with 0.9 billion, 3.7 billion, 7 billion and 32 billion parameters, alongside sparse models with 36 billion and 375 billion total parameters.
The smallest 0.9-billion-parameter model is designed for constrained devices such as smartwatches and glasses, while the 3.7-billion and 7-billion parameter versions target phones and other on-device applications, MBZUAI said.
The 32-billion parameter model is aimed at local hosting and on-premises servers. The 36-billion parameter model activates 4 billion parameters, while the flagship 375-billion parameter model activates 23 billion and is designed for more demanding reasoning, agentic and enterprise workloads.
MBZUAI said the three smallest models achieved state-of-the-art results in their respective size categories across areas including mathematics, reasoning, software engineering and agentic capabilities.
“We believe meaningful AI progress depends on the ability to examine, build upon, and improve the technology, not simply access it through an API,” said Eric Xing, founder of IFM and president and university professor of MBZUAI.
“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.”
The six models share a core architecture, vocabulary, training methodology, interfaces and deployment tooling, with the 0.9-billion parameter model using a smaller vocabulary.
IFM has also introduced a dynamic model-routing technique designed to direct tasks to the most cost-effective model, allowing developers to move between different model sizes as their requirements change.
K2 Horizon uses what IFM calls “diffusion distillation”, a technique that generates blocks of tokens in parallel. MBZUAI said this speeds model performance by roughly three times without reducing response quality.
The models also incorporate a “mixture of value attention” architecture, which the university said improves reasoning without requiring additional computation.
“Rather than release a single model, we’re releasing an entire fleet at once, six models, from one small enough to run on a watch to a flagship built for enterprise reasoning,” said Hector Liu, director of IFM’s Silicon Valley Research Lab.
“Every model is built to compete with the best open models at its size, and everyone ships with the weights, code, training data, and methodology behind it. Developers can prototype on the smallest model, scale to the flagship, and verify every claim we make along the way.”
K2 Horizon follows earlier releases including K2, K2-Think and K2 Think V2 as IFM expands its work on openly available foundation models.
IFM was launched by MBZUAI in May 2025 and operates across Abu Dhabi, Silicon Valley and Paris.
The institute’s wider model portfolio includes the K2 series as well as Jais, an Arabic large language model, and PAN, a world model focused on embodied reasoning and real-world simulation.
MBZUAI is also a strategic knowledge partner in a UAE government programme aimed at training 80,000 federal employees as agentic AI experts. The programme is being delivered through collaboration with national universities and technology companies.
K2 Horizon models are available through Hugging Face and can also be deployed using vLLM and SGLang, MBZUAI said. APIs are available through IFM’s ecosystem of inference partners, including Compass, Cerebras, Amazon Web Services and Nebius.
The models and code are being released under the Apache 2.0 licence.
























