Comment by moffkalast

I think it's also true for many local models. People still use NeMo, QwQ, Llama3 for use cases that fit them despite there being replacements that do better on "benchmarks". Not to mention relics like BERT that are still tuned for classification even today. ML models always have weird behaviours and a successor is unlikely to be better in literally every way, once you have something that works well enough it's hard to upgrade without facing different edge cases.

Inference for new releases is routinely bugged for at least a month or two as well, depending on how active the devs of a specific inference engine are and how much model creators collaborate. Personally, I hate how data from GPT's few week (and arguably somewhat ongoing) sycophancy rampage has leaked into datasets that are used for training local models, making a lot of new LLM releases insufferable to use.