This just isn’t true. 2000s commercial software you buy under a restrictive license, and then extend using an sdk. You can’t modify or distribute the code only your plugin or extension. Open weight models you freely download and modify and extend using open source tools, and create derivative works. The licenses tend to be very permissive (e.g. Apache 2.0 https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE)
Perhaps you are not aware of the thousands of derivative models that have been created based on open weight models? Ones tuned for this or that task, expanded, hacked apart.
> akin to sharing an object file which you can compile into something
I see what you are getting at, but modifying object files isn’t practical or simple or commonly done. A model that’s been post trained isn’t the old model with a lump of new code attached to it. The weights have been irreversibly altered and it’s a new thing now. A derivative work.
My contention is that model weights are a form of source code. They the software AND a modifiable thing, closer to minified js than a binary blob. Less flexible than the full training data would be, in some ways, but more useful in others (unless you have a large data center at your disposal). Would you consider something closed source if it had a bunch of generated code, but didn’t include the code used to generate the generated code? I would.