I like that Meta's models are available for general use under a fairly permissive license, but they've really butchered the word "open".
I think I'm okay with calling Llama "open". It is relatively open compared to, say, GPT-4, a model released by a company with an interesting name.
There are no open source AI models, strictly speaking. Because they are not reproducible even will all the data and training setup. Randomness is embedded in training. It could be done without, but performance suffers. So, only for learning and some academic research.
You can absolutely open source an LLM, it's just that noone has. And you can absouletely bundle the random seem with the training data, it's just that, again, noone has.
And runtime randomness is hardly an issue. We write programs with RNG in them all the time and noone takes issue with that