The feature lets teams stage alternative deployments for an existing Dedicated Endpoint and mirror a sample of production traffic to each candidate. Users can choose an open model from Modal’s catalog or supply weights from Hugging Face or a Modal Volume.
Production continues handling the original requests while candidates receive copies. Each candidate has its own URL for existing quality evaluations; the dashboard exposes inference metrics such as latency and throughput and supports performance benchmarks. A candidate can be promoted from the dashboard, initiating a production rollout at the existing Endpoint URL. Endpoint Candidates use standard Modal compute pricing, including compute for mirrored requests.
Modal says mirroring is best-effort. Its Pingora-based proxy uses bounded per-candidate buffers and a process-level cap on retained request-body data; if pressure exceeds those limits, mirrored requests are canceled while production and other mirrors continue. The announcement describes a workflow and safeguards, not measured improvements in quality, latency, throughput, or cost.