Licenses and what you may do

A model's license is part of its spec, and ignoring it is how people get burned. Open weights do not automatically mean "do anything you like." Before you build a product on a model, read the license.

The common licenses

• Apache 2.0, MIT — permissive. Use, modify, distribute, sell. The closest thing to "no strings."
• Llama Community License — open weights with conditions: usage limits above a scale, and restrictions on using the output to train competing models. Fine for most people, not "fully open."
• Gemma, Qwen, GLM licenses — each has its own terms: some restrict certain commercial use, some require attribution, some add a specific use policy.

There is no single rule. The only safe habit is to read the license for the exact model and version, because terms vary even within a family and change between releases.

The three questions to ask

1. Can I use it commercially? Many licenses permit it, some do not, some permit it only up to a size.
2. Can I ship a modified (fine-tuned) version? Some licenses restrict derived models or require you to open your changes.
3. Are there use restrictions? Some forbid certain domains or require you to follow an acceptable-use policy.

Derived models and fine-tuning

If you fine-tune a model, the base model's license still applies, on top of whatever you add. You cannot take a restrictive base model and release your fine-tune under MIT.

The practical habit

When you pick a model, note the license in the same breath as the VRAM. The catalog records the model family; the license is the thing you verify on the model card before you depend on it for a business.

Being careful here is not legalism, it is the difference between a hobby and a product you can stand behind.