According to their internal benchmarks that don't bench token output though. Wait for someone to get their hands on one before buying just for local LLM purposes.