I’m considering running a DeepSeek model behind a small internal app for a team, rather than letting everyone launch their own copy. I understand that model size and hardware matter, but I’m less clear on what happens when several people send prompts around the same time.

For a DeepSeek local inference server, how do you evaluate concurrent requests, queueing, and response time under actual team usage? I’d appreciate experiences with different serving setups, especially where a single-user test looked fine but shared use changed things.