AIThe day
“Open” models: what the word means, and what it doesn’t
Published weights, source code, training data: in AI, openness comes à la carte. Better to know which kind you are being sold.
Announcements of new “open” models keep coming. The word covers very different realities.
Most often, it means an open-weights model: you can download it, run it on your own servers and adapt it. But the training data and the full recipe stay private, and the licence may restrict some commercial uses.
A strictly open-source model would also publish what you need to reproduce it: code, method and, ideally, data. That is far rarer.
Why it matters: for a company, the difference has very concrete consequences. In-house hosting, data confidentiality, costs, rights of use: before choosing, reading the licence beats reading the press release.