Comment by justaboutanyone

You can run large-ish MoE model at good speeds, like gpt-oss-120b, it's snappy enough even with big context.

But large and dense at the same time is a bit slow.

Running a local LLM will be a load of money for something much slower than the api providers though.