Upgrading AI Server

AI Server has four things that update, on their own schedules.

The Windows app and the containers are numbered separately; the release notes list what each version changed.

Windows app and service

Microsoft Store updates the app in the background. The Windows service runs its own copy of the server, so after an update the app shows that the service is out of date and offers Update service; Windows asks for administrator approval once. A startup task also refreshes the service copy after an update, keeping the configured port. Settings, keys, licence and models are untouched.

Containers

1. Read the release notes for anything marked as needing action.
2. Pull the new tag: docker pull softwaretailor/aiserver:<version> (and aigateway, aiserver-pro).
3. Recreate the container with the same /data volume.

On SIGTERM the old container stops accepting new work (/readyz returns 503), finishes what is in flight, hands its licence seat back, and exits. Give it at least 40 seconds (stop_grace_period: 40s).

A farm without downtime

Behind a gateway, upgrade one worker at a time:

1. Stop the worker. The gateway sees /readyz return 503 and routes around it; in-flight answers finish.
2. Start it on the new version with the same volume, so it keeps its models.
3. Wait until the dashboard shows it healthy, then do the next one.

On Kubernetes helm upgrade --set worker.image.tag=<version> does exactly this with a rolling update. Upgrade the gateway last; with two gateway replicas the Service always has one ready.

To try a new model or version on part of the traffic first, mark one worker as a canary — see scaling and high availability.

Local AI engines

The local engines (the bundled Local AI Engine and the optional llama.cpp engine) are updated from their upstream releases. A new release is downloaded in the background and applied the next time the engine starts, never while it is serving. Turn Update automatically off per engine under Settings → AI engines, or use Apply now to apply a downloaded update straight away.

Breaking changes we have announced

• The legacy local-AI API on port 11434 stops on 2026-12-31. Move integrations to the /v1 API before then — see compatibility.
• After a licence signing-key rotation, older images cannot verify new leases. Keep images reasonably current; the release notes say when a rotation happened.

Questions