What open-source means for frontier models

The phrase open-source is doing a lot of work in AI discussions. Traditional open-source software requires access to source code under a license that permits use, study, modification, and redistribution. Applied to models, that definition fits poorly. A neural network's behavior emerges from weights, training data, training code, evaluation code, and a chain of decisions that are not all equally shareable. Different releases share different subsets, and call the result open-source regardless.

The Open Source Initiative has been developing a definition of open-source AI to distinguish fully open releases from partial ones. Under their approach, a fully open model would make available the weights, the training and inference code, and sufficient information about training data to allow a skilled recipient to recreate a substantially equivalent system. Many widely used releases meet some of these conditions but not others. Weights-available models, sometimes called open-weight, are a common middle category.

The policy debate around openness turns on two concerns. The first is safety, specifically the worry that fully open frontier models could lower the cost of misuse in areas like cybersecurity and biosecurity. The second is competition, specifically the concern that closed frontier development concentrates power in a small number of firms. The US National Telecommunications and Information Administration completed a report on open-weight models in 2024 that treated both concerns seriously without recommending an immediate policy change.

Editor's note: Using open-source as a single label hides the question that actually matters for any given model, which is what a recipient can do with the release and what accountability travels with it. Journalists, procurement teams, and policy writers would communicate more clearly by naming the specific artifacts a release includes. The trend toward more precise vocabulary, open-weight, open-data, open-training, is worth supporting.