Is the list open? I would probably pick among some local models and ones from nano-gpt myself if I had the option. And is the criteria for type of model needed code reasonably accessible? It would be an interesting thing to hack and maybe train.