I revisited this and using the models directly via ollama run was actually surprisingly fast.

A bug or misconfiguration with the connection to opencode seemed to be the culprit.