The model is a process, not a vendor

Your product is a small model plus an API plus some state. On hosted-inference platforms, the model lives on someone else's fleet, and every request your API serves detours across the internet and back.

On Hoody, the whole product is one container. The model server: sized for the box it runs on: is a daemon program with its weights on the container's own disk. The API in front of it is an exec script calling it over localhost. The job queue and results are a SQLite database on the same filesystem.

One machine, three processes, zero network hops between your API and your model. Deploying a new version means restarting a daemon, not coordinating with a fleet you don't operate.