The "World's Largest Fully Open AI" Just Launched — But Is It Actually Fully Open?

MBZUAI's Institute of Foundation Models (IFM) is calling K2 Horizon "the world's largest fully open AI models in history" — a six-model lineup ranging from 0.9B to 375B parameters. What sets this apart from a typical open-weights release is the promised scope: final weights, training code, intermediate checkpoints, training logs, and evaluation results, all meant to let outside researchers inspect and reproduce the training process — not just run the finished model.

The 375B flagship is the headline number, but the real test is the openness claim itself. "Fully open" is doing a lot of work here, and the announcement leaves gaps. MBZUAI says the release includes training data; IFM's own blog says it's "data or data-construction recipes" — which are not the same thing. Having the actual dataset versus instructions to rebuild one are very different levels of transparency, and it's unclear which applies to which of the six models.

Licensing clarity is also thin. Apache 2.0 is named as covering the release, but it's not specified whether that applies uniformly across weights, code, datasets, and checkpoints — a real question for any team considering commercial use or redistribution.

Bottom line: K2 Horizon is a genuinely ambitious open-AI commitment on paper — bigger in scope than most "open" releases that just drop a downloadable model. Whether it earns the "fully open" label depends on what's actually published, under what terms, and whether outside developers can use it to meaningfully reproduce the work. That verification hasn't happened yet.