June 7, 2026
Dedicated is renamed to Serving; API keys, webhooks, and base-model completions launch; live performance metrics and clearer failure diagnostics.
Major features
Section titled “Major features””Dedicated” is now “Serving”
Section titled “”Dedicated” is now “Serving””The Dedicated section of the console has been renamed to Serving — new sidebar icon, routes, and terminology. This is a naming change only; nothing about how your deployments behave has changed.
API keys, webhooks, and base-model completions
Section titled “API keys, webhooks, and base-model completions”- Added an API Keys tab on the account page to create and manage keys.
- Added webhook subscriptions with delivery visibility and test-send, so you can get notified when async work finishes.
- Added
/v1/completionssupport and a matching Playground mode for base (non-chat) models.
Improvements
Section titled “Improvements”- Serving deployment cards now show live performance metrics (throughput, GPU utilization, and more) instead of just a status badge.
- New deployments show weight-download progress (percent and bytes) instead of a generic “starting” state.
- Deployment failures now show a specific out-of-memory diagnosis (GPU memory vs host memory) instead of a generic error, and a banner appears when scaling up is blocked by a GPU shortage.
- More accurate memory sizing for image/video model deployments, reducing spurious out-of-memory failures.
- Serving replicas now self-heal: unreachable replicas are automatically detected and replaced.
- Official models can now be scheduled for deferred retirement, with a warning banner on affected deployments.
- Port checks during host onboarding now retry automatically after a short settle window, in addition to the manual retry button.
- Fixed intermittent provisioning and auto-scaling failures for Serving deployments.
- Synchronous inference errors now show a friendlier message that hints at possible out-of-memory causes.