Why aren't languages/frameworks offering retrained models for their project?

[ckluis]

The cost of training is coming down.

We have incredible open source models (especially smaller ones) like qwen 3.6-27b.

Imagine the more niche languages/frameworks getting a model rebuilt for them: zig, haskell, elixir/ash, f#, IHP, jetzig, etc.

I could imagine that there would be tremendous value in a language + framework model like: zig + jetzig that is explicitly trained on the materials and best practices for these.

Which begs the question - at what point does it make sense to make a model specifically for narrower and narrower contexts. How much faster/better could they become?

[mmarian]

I remember seeing somewhere that LLMs performed on a niche subset of data perform worse than general ones. RAG is probably best here.