Per-model configurable max_tokens?

[mdschmitt]

Pretty much the subject. I'd love to be able to set Haiku to 4096 max_tokens while letting it be bigger for beefier models.