Jargon:Open-weight model
1 min read
Say this at the party
“An AI you can download and run yourself. You get the finished brain, but not usually the recipe or the data it learned from.”
In plain English
An AI model's "weights" are the billions of numbers it learned during training. They are, in effect, the model.
An open-weight model is one where the company lets you download those numbers. You can run it on your own computer, tweak it, and build things with it, often for free.
Open source is a bigger promise. It means you also get what you need to rebuild it: the training code and enough detail about the data. Most "open" AI models give you the weights only.
Where it came from
Open-weight models took off in 2023, when Meta released Llama 1 and 2 and Mistral AI released its own models. In October 2024, the Open Source Initiative published its Open Source AI Definition, which says weights alone don't count as open source.
See it in action
A small law firm downloads an open-weight model and runs it on its own server. Client files never leave the building, and there's no monthly bill per question.
What people get wrong
- Open-weight isn't the same as open source. Without the training code and data details, you can use the model but can't truly see how it was made or rebuild it.
- It isn't always free to use however you like. Some come with licences that limit big companies or certain uses, so read the small print.
Go deeper
- ArticleWikipediaOpen-source artificial intelligenceExplains the open-weight versus open-source debate and "openwashing".
- GuideNTIA, US Department of Commerce, 2024Dual-Use Foundation Models with Widely Available Model Weights ReportA government look at the benefits and risks of releasing model weights.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.