“Open source AI model” gets used to describe releases that differ enormously in what’s actually made available, and the gap between the most and least open releases labeled this way is large enough that the term has arguably stopped conveying reliable information on its own without checking the specific license and what was actually published.
The Four Things That Can Be “Open”
A model release can independently open up its weights (the trained parameters, letting anyone run the model), its training code (how the model was built and trained, letting anyone reproduce the training process), its training data (what the model actually learned from, letting anyone audit or reproduce the dataset), and its license terms (whether commercial use, modification, and redistribution are genuinely unrestricted or carry meaningful conditions). Most releases marketed as “open source” open only the first of these — weights — while keeping training code and data proprietary, which is a materially narrower and less reproducible kind of openness than the term implies to anyone assuming full-stack transparency.
Why Weights-Only Release Isn’t Full Transparency
Having a model’s weights lets you run it, fine-tune it, and build on top of it — genuinely useful capabilities — but it tells you very little about what data shaped its behavior, what biases or limitations that data introduced, or how reproducible the training process actually is if you wanted to build a comparable model independently. This is a meaningfully different kind of openness than open-source software has traditionally meant, where source code availability generally does let you understand and reproduce how the software actually works from first principles.
The License Fine Print That Actually Matters
Several prominent “open” model releases carry license terms with real restrictions — usage caps tied to a company’s revenue or user count above which a separate commercial license becomes required, or restrictions on using the model’s own outputs to train a competing model — that meaningfully diverge from the unrestricted use the Open Source Initiative’s formal definition requires, and from what most people assume “open source” guarantees when they see the label applied. Reading the actual license before building a product around a specific model’s availability is a genuinely necessary step, not excessive caution, given how much variation exists behind an identical marketing label.
Why the Label Still Matters Despite the Confusion
Even weights-only, restrictively-licensed releases represent real, meaningful movement compared to fully closed, API-only models, because they still let researchers and smaller companies run, inspect the observable outputs of, and build directly on top of the model — genuinely valuable capability closed models never offer regardless of the license’s specific restrictions. The practical, useful move isn’t dismissing “open” as a meaningless marketing term, but reading past the label to what was actually released and under what specific terms before making a build decision that depends on it.