[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.

[storystarling]

Makes sense regarding the MoE performance. I am not sure the cost argument holds up for high volume workloads though. If you are running batch jobs 24/7 the hardware pays for itself in a few months compared to API opex. It really just comes down to utilization.

[storystarling]

Do you have specific t/s numbers for those dense models? I'm curious just how severe the memory bandwidth bottleneck gets in practice.

I'm not sure I agree on the cost aspect though. For high-volume production workloads the API bills scale linearly and can get painful fast. If you can amortize the hardware over a year and keep the data local for privacy, the math often works out in favor of self-hosting.

[justaboutanyone]

For Qwen2.5-72B-Instruct-Q5_K_M at 32k context, I fed it a 26k token file (truncated fiction novel) asking it to summarize, and it input processed at 224 tok/s and output generated at 3 tok/s. Not really good enough for interactive use without frustration. Not just from watching it reply, but also the long wait for it to actually read the book.

On the same hardware gpt-oss-120b at 128k context, I fed it a longer version of the input (a whole novel, 97k tok), and it input processed at 1650 tok/s and output generated at 27 tok/s. Just fast enough IMO