Model serving
Aidin serves registered models on managed inference endpoints built on KServe. The platform runs the predictor, routes traffic to it, and exposes it to your workloads through the platform’s standard access layer.
What is available today
Section titled “What is available today”| Capability | Availability |
|---|---|
| Managed inference endpoints for deployed model artifacts | Available |
| Standard predictor runtimes (for example vLLM, Triton, TorchServe class runtimes) | Available |
| High-throughput LLM serving engine (private vLLM-based) | Coming soon |
| Serverless scale-to-zero endpoints | Coming soon |
Availability reflects the platform feature catalog state (2026-11-10). Endpoints marked coming soon are not yet offered — they are not documented as how-tos anywhere on this site.
How serving relates to other pieces
Section titled “How serving relates to other pieces”- Models come from the model registry.
- Training happens in notebooks, pipelines, or Katib experiments.
- Endpoint compute is subject to the same entitlement and quota model as other workloads: see Cloud resources.
Upstream reference
Section titled “Upstream reference”KServe is an open-source project; Aidin runs a managed, version-pinned deployment of it. Where Aidin’s behavior differs from upstream KServe documentation, Aidin’s documented behavior takes precedence.