Lexicon · I · The model
Open weight
Also: open-weight model
A model whose weights are published: anyone can download it and run it on their own machines, without ever holding the full recipe of how it was made.
The analogy
A great chef delivers you their signature dish, vacuum-sealed. You can reheat it in your own kitchen, serve it to your guests, taste it as often as you like. What you do not have is the recipe: the exact ingredients, the proportions, the technique. Open source is the published recipe; open weight is the delivered dish.
A concrete example
Llama (Meta), the Mistral models or DeepSeek are open weight: a bank can host them in its own datacenter so customer data never leaves its walls. It is the typical choice of regulated industries (healthcare, defense, finance), where "where does our data go?" outranks raw performance.
What it is NOT
It is not a guarantee of full transparency: the training data remains unknown, so nobody knows what the model has "read". It is not a security or quality label. And it is not the opposite of seriousness: closed and open-weight models coexist at the highest level: it is a distribution choice, not a hierarchy.
Often confused with
The distinction that changes the conversation.
- Open source
- Open source publishes the recipe: code, ingredients, method. Anyone can rebuild and modify the software. Open weight publishes only the result: the weights. You can use the model, not reproduce it: the training data and process stay secret.
- Free
- Downloading the weights costs nothing, but running them requires expensive machines, and the license may restrict commercial use. Open weight describes a publication mode, not a business model.