Licensing — open source vs. open weights

I'll admit this openly: the commercial implications of this distinction only really clicked for me with Flux.2 dev, after I'd already been using the model. "Open weights" sounds like free use — the model is, after all, publicly downloadable, anyone can inspect it, anyone can run it themselves. That feels like open source. It isn't necessarily, though.

This mix-up isn't embarrassing — it's an easy one to make. That's exactly why it gets its own article here, before it turns into an expensive misunderstanding.

With genuine open source, it's not just the result that's publicly visible — usage itself is permissively licensed: download, modify, redistribute, use commercially, all without special permission, without an additional license fee, without restriction to particular use cases. SDXL is an example of this: its license permits commercial use without a separate agreement.

That's the expectation the word "open" creates in everyday use. And that expectation isn't met by many current AI models — even though they describe themselves as "open" too.

Open weights initially means only this: the model's trained weights — the file, the checkpoint from the previous articles — are publicly accessible. You can download them, run them locally, build them into your own workflows. What you're allowed to do with the outputs (the generated images) and what you're allowed to do with the model itself are two separate questions — and that separation is exactly the point that's easy to overlook.

With Flux.1 dev and Flux.2 dev from Black Forest Labs, for example: the generated images may be used for private, scientific, and commercial purposes. The model itself — the weights, the checkpoint — is under a non-commercial license, though. Anyone wanting to run or redistribute the model in a commercial context needs a separate, paid commercial license from the manufacturer.

With Ideogram 4 the structure is similar, just drawn differently: the openly available weights are intended for research, evaluation, and personal projects — non-commercial. Anyone wanting to deploy the model in production needs a commercial license, which then also grants access to the full model version (the open variant is additionally available only at reduced precision).

The practical consequence: "I can download the file and it runs on my machine" says nothing about the license. It's only a statement about technical availability. Whether you're actually allowed to make money from what you produce with it is governed by an entirely different document — the license agreement that has to be accepted on download, but is rarely read in full.

The current generation of image models mostly operates under this open-weights model, not the classic open-source model. That's a deliberate business strategy on the part of manufacturers: openness for research, visibility, and community building, while monetizing commercial use through separate licenses. Understandable from a business perspective, but not open source in the original sense.

For you as a user, that means: with every new model, it's worth taking a second look at the actual license, not just the "open" marketing label. The terms worth watching for: non-commercial license (use of the model itself is restricted to non-commercial purposes), commercial license required (a separate, often paid agreement is needed), and explicit clarification of whether a permission applies to the output (the generated image) or to the model (the weights, the software) — these are two different things, even when one license document covers both.

RAY-L is a bridge between Blender and ComfyUI, not a model provider. Which model you use within RAY-L — and under what license you're allowed to use it for that purpose — is your own responsibility as the user. Even though RAY-L itself works smoothly with every supported model, you're the one who ultimately checks whether that model's license permits your intended use, including any commercial use.