These models use the completions endpoint, not chat/completions, and return 400 errors when used via the chat API.