How about the freedom to choose a provider for the model? Like maybe it's determined that gpt-oss-20b is good, but the user gets to decide on provider and whether it can be local (which would entail a quantized version of the model so technically a different model?)?
