I’ve mostly used Hugging Face for model discovery and quick experimentation.

How others manage moving from experimentation to production — do you stick with hosted inference or switch to custom pipelines?