Speech models now unload after 120 idle seconds by default. Set speech.idle_timeout_seconds in advanced speech settings or the shared configuration interface to a whole number from 1 through 86400. Valid saved changes apply live. Idle time starts after the latest completed use; active transcription retains the model, and the next request safely reloads it.
A bounded Linux amd64/glibc check with a two-second override reduced RSS from about 1119 MiB after transcription to 53 MiB after automatic cleanup; the reload cycle fell from about 1214 MiB to 55 MiB. Reload construction took 2.879 seconds. These warm-cache measurements used explicit model paths on a shared host, bypassing managed-cache provisioning and per-use integrity work. They are not performance guarantees or measurements of other platforms, allocators, cold-cache latency, or long-run memory behavior.
Worker protocol is now 17 so canonical settings retain the timeout field. Before updating an older installation, drain jobs and pending delivery, explicitly stop that installation, then update and relaunch. The existing incompatible-worker refusal remains in place.