AI Explained
Open-weight vs open-source AI models
Lots of AI models are called open, but most aren't open source. What open weights give you, what the OSI's Open Source AI Definition requires, and how to check.
On this page
Lots of AI models are called "open". Most of them aren't open source. The difference matters when you build on them.
Open weights: the cake, not the recipe
An open-weight model gives you the trained weights: the file you download and run. But you usually don't get the training data or the training code. It's like getting a finished cake without the recipe.
What open source AI actually means
In October 2024, the Open Source Initiative published the Open Source AI Definition 1.0. It requires four freedoms, for any purpose:
- Use the system.
- Study how it works.
- Modify it.
- Share it.
And it requires three parts to be available: the parameters (weights), the complete code to train and run the model, and enough detail about the training data that a skilled person could build a substantially equivalent system.
Three real examples
Meta's Llama 3.1: open weights with rules. You can download the weights, but the license adds conditions. Companies with over 700 million monthly active users must request a separate license from Meta, and everyone must follow Meta's acceptable use policy. That's open weight, not open source.
Mistral 7B: a standard open license. Mistral released it under Apache 2.0, which it said "can be used without restrictions". Still, the training data isn't released.
AI2's OLMo: fully open. It releases the weights, the training data and the code.
Closed models give you none of that. You send requests to an API and get answers back.
Think of it as a spectrum
Closed API → open weights with a custom license → open weights with a standard license → fully open.
Why it matters
- Open weights let you run a model yourself, offline, and fine-tune it.
- Full openness also lets you check how it was made, and reproduce it.
Before you ship, read the license
Look for:
- limits on the number of users,
- banned uses,
- rules on using the model's outputs to train other models.
Don't trust the word "open" on its own. Check what's actually released.
Open weights aren't open source.