Model pages now show cold or warm state

Model pages now show whether the version you're viewing is currently cold or warm. Warm means at least one instance is active, and requests are served with minimal latency. Cold means no instance is running, and the next request incurs startup time while the model is loaded.

This is per version, not per model. A model with multiple versions can have one version warm and another cold at the same time, since WriftAI scales each version's instances independently based on its own traffic.

A screenshot showing hexgrad/kokoro-82m in cold state

A screenshot showing hexgrad/kokoro-82m in warm state

What you can do with this

Latency varies between a cold and a warm run, and until now there was no way to check which one you were about to get. If you're benchmarking latency or debugging a slow request, you can check state before you run anything instead of inferring it from timing after the fact.

Cold and warm aren't permanent. WriftAI scales instances out with traffic and scales them down toward zero when idle, so a version with no recent requests eventually goes cold. Sending it a request brings it back up. Read more in how WriftAI works.