Open Source vs Closed AI Models

"Open source AI" gets used loosely, and it's worth being precise about what it actually means for a model — because it's genuinely different from what "open source" has always meant for regular software, and the difference matters for anyone deciding which kind of model to build on.

What "open" actually means here

For traditional software, open source means the source code— the actual instructions — is published, so anyone can read, modify, and rebuild it. For an AI model, the equivalent artifact isn't code in that sense; it's the weights — the finished, trained set of parameters covered two lessons ago. Publishing the weights lets anyone download and run the model themselves. It usually does not mean the training data or the exact training process is public — those generally stay private even for models widely described as "open."

That distinction is why many people in the field now prefer the more precise term open-weight over "open source" for models — it says exactly what's actually open, without implying a level of transparency (the training data, the process) that usually isn't there.

Real examples

Meta's Llama family, Mistral AI's models, and DeepSeek's releases are the best-known examples of labs regularly publishing open weights — you can download them and run them on your own hardware. That sits opposite OpenAI's GPT models, Anthropic's Claude, and Google's Gemini, which are only accessible through an API call to the provider's own servers — no weights to download, no way to run them yourself.

Exact licensing terms differ across open-weight releases and change over time, so it's worth reading the specific license for any model before building a product on it — "open-weight" doesn't automatically mean "free to use however you want commercially."

Why companies release them at all

A company giving away something that cost enormous amounts of compute to train looks counterintuitive until you consider what it buys them: developer mindshare and an ecosystem built on their model, competitive pressure against closed leaders, and access to a wider community finding bugs and building improvements they wouldn't have found alone. These aren't the only reasons, and different labs clearly weigh them differently — but they're real, defensible business reasons, not pure altruism.

The real tradeoffs

Toggle between the two to see what you're actually trading off.