Open Weights and Open Source Sound the Same. One of Them Can Get You Fired.
When a model is 'open weight' you can run it, but you often can't see how it was made or ship it freely. That gap matters if you work anywhere with a legal team.
I used to say 'open source AI' and mean three different things. They are not the same, and mixing them up in a meeting is how you promise something your compliance person will block.
Open source means the code is published and you can inspect and modify it under a license. Open weights means the trained model file is published, but you may not see the training data, the code, or have permission to use it commercially. One is a book with the recipe; the other is the finished cake with the oven still locked.
Why it bites: a weight you downloaded 'for free' can carry a license that forbids commercial use or requires you to publish your own changes. The model runs on your laptop and still lands you in trouble if the terms say so.
The practical rule I use: if I'm building a product, I read the license before I download, not after. 'Open' is a feeling; the license is the contract.
Ollama (ollama.com)
Best for: developers who want to run open-weight models on their own machine
Key features: one-command model pulling, local inference, API compatible with common tools
Free plan: free and open source to use
Pricing: free; you pay for the hardware and electricity
Pros: private by default, no data leaves your machine, great for experiments
Cons: you need a decent GPU for big models, and model quality varies by release
Open weights gave me a lab on my desk. Open source gave me the right to understand it. Knowing which is which is the difference between a cool demo and a defensible product.
Frequently asked questions:
Q: Is an open-weight model free to use commercially?
A: Not always. Check the license; many forbid commercial use or require disclosure.
Q: Why run a model locally at all?
A: Privacy, cost at scale, and no dependency on someone else's API going down.
Q: Does open source mean I can see how it was trained?
A: Often not. Weights are the result; the training code and data are separate questions.