The Institute of Foundation Models (IFM) at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) released K2 Horizon, a fleet of six fully open AI foundation models, ranging from 0.9 billion to 375 billion parameters.
The launch sits on the backdrop of a global debate on “open” AI models, which has become louder in the past two months. As Chinese open-weight models close the gap on frontier models, reports circulated that the Trump administration could consider restrictions on advanced Chinese models.
More than 270 companies signed a letter urging Washington not to restrict open-weight models prematurely, while others (most notably Anthropic CEO Dario Amodei) cautioned that open weights are harder to monitor and safeguard, even while acknowledging their value for research and competition.
Amid the noise, Stanford HAI argued that the debate is centred on the wrong topic – open-weight models are not enough, and science and society actually need fully open models.
"Open weights allow people to download and run a model, but they do not necessarily allow researchers to understand what data shaped its behaviour or reproduce how it was trained,” Richard Morton, vice president and executive director of the IFM, tells WIRED Middle East.
Open-weight models allow access to the artifact, but fully open models allow anyone to access, inspect or adapt weights, code, training data and methodology.
“This matters as AI increasingly moves into areas such as government services, healthcare, education and critical infrastructure,” Morton says. “If we are going to rely on increasingly capable systems, we should also be able to examine, understand and improve them.”
The K2 Horizon fleet includes 0.9B, 3.7B and 7B models – the former designed to work in constrained environments like glasses or watches, and the latter two developed for phones. The dense 32B and sparse 36B-A4B models provide options for local hosting and on-premise servers, while the 375B-A23B model is designed for enterprise deployments.
Morton said that its smallest models are “state-of-the-art”. He says the 0.9B is the industry's best performing of its size in mathematics, reasoning and tool use; the 3.7B has the best reasoning performance under 4B parameters and “can match or exceed many larger models”; and the 7B is the industry’s best-performing model under 10B parameters.
He adds that the 36B-A4B uses mixture of value attention architecture to “outperform many larger models while activating only 4B parameters”, and the 375B-A23B flagship competes with leading open-weight models on reasoning and agentic work.
“So, across the fleet, we are looking not only at raw parameter count, but at how much capability each model can deliver at its particular scale and computational footprint,” he says.
IFM’s dynamic model routing sends tasks to the most cost-effective model, a practical measure for developers using it from prototyping to production. K2 Horizon’s diffusion distillation technique generates blocks of tokens in parallel, which it says speeds model performance three-fold without degrading the quality of the response.
“Developers can prototype on the smallest model, scale to the flagship, and verify every claim we make along the way,” said Hector Liu, director of IFM’s Silicon Valley Research Lab, referring to the decision to release six fully open models at once.
“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 Professor 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.”
