Expose one virtual model name and route it to provider targets or hardware-aware local Ollama model selectors.