You can still leverage knowledge from a foundation model in a smaller fine-tuned one.

So the model might have learned general OOP principles from Java but it then drops redundant parameters about specific conventions like AbstractFactoryBuilders when it specialises on a language like Python which has no notion of Interfaces.

Likewise real world knowledge might help distinguish between accounting and database transactions when writing a banking application but you don't necessarily need your coding assistant to have memorised all the world cup winners since 1966.

These models are unwieldy so I think it makes a lot of sense to try and find ones that are tuned efficiently.

I’d say it depends. Some patterns will be different. Like trying to create zig code when the LLM is trained mostly on javascript and python.