Your point about caliber/quality is fair, but I have been pretty astonished by some of the newer/better models (Gemma 4 variants, GPT-OSS before that).
However, there's not a lot of memory increase to have multiple sessions in parallel with one model. It's an HTTP server, and other than some caching, basically stateless.
However, there's not a lot of memory increase to have multiple sessions in parallel with one model. It's an HTTP server, and other than some caching, basically stateless.