Ask HN: Who is using small OS LLMs in production?

It is not clear whether relatively small open source LLMs in production or not, for instance replit code 3b, llama 2 7b, codegen,... What are the motivations for using those models over prompted GPT api ?