LM Studio models unload after requests

This feels less like “your model is too big” and more like LM Studio’s serving stack tripping over its own unload/load lifecycle. The frustrating part is the silent failure: without logs, every crash looks like a model-vanished problem.

• –LM Studio docs say API-loaded models can auto-unload after an idle TTL, so a disappearing model may be intentional if the server is configured that way.
• –The user is hitting LM Studio on another server, which points the finger at the remote inference process, GPU/driver stability, or request-handling state rather than the local desktop app.
• –LM Studio now leans into headless and remote deployments, so API stability under repeated requests is a core infrastructure expectation, not a nice-to-have.
• –If this reproduces on small models, size is probably not the root cause; config, evict behavior, or backend crashes are more likely culprits.