Abu Dhabi University Releases Massive Open-Source AI Models, Prioritizes Transparency Over

MBZUAI’s Institute of Foundational Models released K2 Horizon on Thursday, a suite of six artificial intelligence models that the Abu Dhabi-based university describes as among the largest fully open systems in the field’s history. All underlying weights, code, training data and methodology are made publicly accessible, rather than restricted behind proprietary application programming interfaces.

Prof Eric Xing, president and professor at MBZUAI, framed the release as a deliberate departure from prevailing industry practice. “We believe meaningful AI progress depends on the ability to examine, build upon and improve the technology, not simply access it through an API,” he said. “The most important technology of our time should be built with the world, not kept from it.”

The suite spans a wide range of computational scales. The smallest model, designated 0.9B, contains 0.9 billion parameters and is engineered to operate on devices as constrained as wristwatches. At the other end sits the 375B model, designed to compete with leading open-weight systems on reasoning and agentic tasks. In AI terminology, parameters are variables that enhance the accuracy of large language model predictions.

Hector Liu, director of IFM’s Silicon Valley research lab, stressed the verifiability built into the architecture. “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,” he said. “Developers can prototype on the smallest model, scale to the flagship and verify every claim we make along the way.”

Two technical innovations underpin the suite. K2 Horizon uses what MBZUAI calls “dynamic model routing,” which directs specific computational tasks to whichever model in the suite offers the most cost-effective execution. It also employs a “mixture of value architecture” designed to enhance reasoning capabilities without requiring proportional increases in computational resources.

By contrast with closed systems that gate access through APIs, the models are available for download across Hugging Face, Ollama, Unsloth and the IFM Repository, allowing researchers and developers globally to inspect and modify the systems directly.

MBZUAI was established in 2020 as the world’s first university dedicated exclusively to artificial intelligence research and education. The K2 Horizon release sits within a broader national strategy: the UAE has positioned itself as a developing AI hub, diversifying its economy away from hydrocarbon dependence through targeted investments and partnerships. The country has partnered with the United States to construct an AI campus featuring 5 gigawatts of data centre capacity in Abu Dhabi, and recent years have brought collaborations with Microsoft, Nvidia and OpenAI alongside the emergence of AI-focused start-ups.

Whether the fully open approach adopted by MBZUAI will influence how other institutions and governments structure their own AI development commitments remains an open question, particularly as debates over transparency and national competitiveness in the sector intensify.