Question on Achieving Real-Time Performance in M3-Agent

Hi, thank you for sharing the great work on M3-Agent.
I have one question regarding the implementation details.

Could you kindly explain how the system achieves real-time performance in practice?
I’m especially interested in whether this was obtained through model optimization, system-level parallelization, or specific engineering tricks.

Additionally, it would be very helpful to know:
- How you measured "real-time" performance (e.g., latency per frame, end-to-end response time).
- Under what hardware/software environment the reported performance was observed.

Any clarification would be greatly appreciated.
Thanks again for your excellent contribution! 🙏