Local inference is already very good on open models if you have the hardware for it.