How's the performance compared to ordinary 9B LLM with structured outputs? Both accuracy and speed?