Indeed, but it probably won‘t get any easier for us devs when clients also get easier access to open weight LLMs. I mean from personal experience 40 tok/s on an M3 pro with gpt-oss-20b holds up quite well for lots of tasks. Thinks are changing so fast.

[dead]