Forty fine-tuned models, forty containers, one of them awake

You've fine-tuned forty variants of a small model: per client, per domain, per experiment you didn't want to lose. Serving all forty means forty resident model servers, and a resident model server wants gigabytes of RAM whether anyone calls it or not. On per-endpoint hosting, forty endpoints is forty bills. You kept six.

On Hoody each variant is a container with its weights on disk and its server registered as a daemon program, stopped. Stopped, it's files: no RAM, no CPU, disk blocks shared copy-on-write with the base-model image the other thirty-nine were cloned from. When a client needs variant 23, you POST start on its daemon, it loads, it serves. When they're done, you POST stop.

The weights never left the box. The box never got bigger.

Forty models used to mean forty servers; now it means forty folders and a start button.