What's the difference between closed, open‑source and open-weight AI? A researcher explains - PBS

What the report says

PBS published an explainer from The Conversation on July 25, 2026, by Jeffrey Young, a principal research scientist at Georgia Tech’s Partnership for an Advanced Computing Environment, outlining how the terms “closed,” “open-source” and “open-weight” are used in discussions about artificial intelligence models. The article says the labels describe how much information a model developer releases about a system’s design and whether others can inspect, alter or reuse it.

According to the piece, open-source software grew out of the free software movement of the 1980s and 1990s, which emphasized users’ ability to run, examine, modify and redistribute programs. In software, that typically depends on access to source code and on licensing terms such as the GNU General Public License, Apache License, MIT License and Berkeley Software Distribution license.

The article explains that AI has complicated the definition. Large language models are trained on large datasets and then released for public or private use. Some companies disclose code and model “weights,” which represent learned patterns from training, while keeping other elements unavailable. PBS/The Conversation cites Meta’s LLaMA release in 2023 as an example that included inference code and weights, but notes that open-source advocates have questioned whether its license qualifies because of limits on commercial reuse.

The explainer also identifies “open-weight” models, including DeepSeek and Alibaba’s Qwen, as systems whose weights are available under comparatively less restrictive terms. The article says some developers and the Open Source Initiative argue that a fully open-source AI model should also include training data, though the size and complexity of those datasets make distribution difficult. The distinction matters because openness affects transparency, reuse, competition and independent scrutiny of AI systems.