"Open source" has a precise, decades-old meaning in software: the source code is public, and you can read, modify, and redistribute it. Applied to AI models, the term gets used far more loosely — and the gap between the label and what's actually shared matters if you're choosing a model based on it.
What usually gets released
When a company calls a model "open source," what's typically public is the trained weights — the billions of numbers the model uses to generate output — packaged so anyone can download and run them. That's genuinely useful: you can self-host, fine-tune, and inspect how the model behaves.
What usually doesn't get released
The training data almost never comes with it. Neither, in most cases, does the full training code or a detailed account of how the model was tuned afterward. That means you can run the model, but you generally can't reproduce it, verify exactly what it was trained on, or audit it the way you could audit open source code line by line. Some researchers use the more precise term "open weight" for exactly this reason.
The license adds another wrinkle
Even models that release weights openly often attach licenses with real restrictions — caps on commercial use above a certain size, restrictions on using the model's output to train competing models, or approval requirements past a certain scale. "Open" and "free to use however you want" aren't the same claim, and the fine print is where they diverge.
Why the distinction matters when choosing a provider
None of this makes open-weight models worse — it just means the label is doing less work than it sounds like. What actually matters for a production use case is usually more practical than the licensing philosophy: reliability, cost, latency, and — for anything answering questions from your own content — how well the model performs once it's grounded in your actual documents, which levels a lot of the differences between models to begin with.
Related reading
- Why two AI models can read the same prompt and disagree — how training differences between any two models, open or closed, show up in practice.
- What is RAG, explained for non-engineers — why grounding matters more than which base model you start with.
- What does it actually cost to run an AI chatbot? — where model choice shows up on a bill.