The reason is because "open source" is a vague enough concept anyone can project what they want onto it. Being able to fully replicate the construction of the model is one lens on it, but being able to modify it to suit your own needs is another.

Given that a released base model allows things like fine tuning, generating your own training sets, and so on, it's not obviously closed source in the way the poster is acting like it is. Releasing the full training data and reproduction pipeline would be "more" open source, but having the raw, non RLHF'ed model is extremely open compared to most alternatives beyond that.

It's not as vague as you're making it sound.

Open source would mean what the engineers who work on this build artifact edit. This is clearly not that.

This is 'more open' than other releases, yes. That doesn't mean it reaches the bar of what 'open source' means.

If a company makes a product that involves a code generator, and publishes the generated code, but not the code generator itself, it’s still open source.

"Some people are right but if you tilt your head and squint you can almost understand the argument from the others"

I don't know how to respond to these lines of reasoning anymore except to say that words mean things and it's worth keeping definitions consistent for the sake of our common understanding.

Machine learning models force us to reconsider the definitions or introduce new ones. A published model has characteristics of open source and closed source software, given the extent to which it can be fundamentally modified, inspected, and used to generate derivative works in a way executable binaries cannot.