A closed model can only be used through an interface (API): you send a text and receive a response, with no access to its insides; GPT-4 and Claude are examples. An open-weights model, by contrast, publishes its parameters and architecture: it can be run on one's own computer, fine-tuned and studied (Alammar & Grootendorst, 2024, pp. 29–30). The LLaMA family was the emblematic case: models from 7 to 65 billion parameters, trained exclusively on publicly available data and released to the research community (Touvron, Lavril, et al., 2023); the weights, meant for restricted access, leaked within days. With quantization and efficient fine-tuning, many of these models now run on consumer hardware.
"Open" is ambiguous, and the ambiguity matters. Publishing the weights is not publishing the data or the training code, and many licenses restrict commercial use; Alammar & Grootendorst (2024) note that this is why what counts as "open source" is disputed (2024, p. 30). The Open Source Initiative (2024) requires three things before calling AI open source: sufficient information about the data, the complete code, and the parameters. A useful analogy: weights without data resemble a compiled program without its source code. You can run it and tweak it, but not rebuild it.
For research, the difference is epistemic. Millière & Buckner (2024b) warn that many striking results about closed models are anecdotes that are irreproducible by design, since no outsider can intervene on them, and that academic research (including interpretability work) has mostly used accessible models such as BERT, to the point that the field was nicknamed "BERTology" (2024b, sec. 3.3). In the safety debate, open models worry people because they cannot be "recalled"; Santelli (2026) points out that the cat is already out of the bag: there are ever more "gardeners" able to grow powerful models with fewer resources.