A small model has intermittent traffic and can tolerate startup latency. The team wants to avoid managing a continuously running instance fleet. Which option should be evaluated if its constraints fit?

1. A batch-only job while promising synchronous submillisecond startup.
2. A training cluster kept running as an undocumented HTTP server.
3. SageMaker Serverless Inference.
4. The largest dedicated endpoint permanently, regardless of utilization.