Traceable rollouts and a new evals surface.
New
- Every rollout now streams to
/tracesin real time so you can watch a deploy land end to end. - New
evalsresource for scheduling and inspecting evaluation runs from the API. - Multi-region routing exposed on
POST /deployments. Requests balance across the regions you name.
Changed
- The default
maxreplica count for hosted models moved from 2 to 4. - Trace retention on free workspaces raised from 7 days to 30 days.
Fixed
- Rolling back a failed deploy now completes in under 5 seconds, down from a reported 40 seconds.