Superlinked is launching SIE Serverless, a hosted inference API for small open models on workloads that spike and then sit idle. It is for software teams that embed and rerank search results, run OCR and document extraction, classify requests, and make other narrow agent calls a focused model can handle. New accounts get $20 in credits with no card required, and teams can apply for another $500 in launch credits. The hosted API runs the same Apache 2.0 Superlinked Inference Engine a team can deploy itself, with one model catalog and one request format on both sides. Embedding, reranking, extraction, and generation models load onto shared GPUs as requests arrive and unload when memory runs short, so each model does not need its own always-on server. Moving a sensitive workload into a team's own cloud on AWS, GCP, Azure, or Alibaba Cloud means changing the address, not the client code. Superlinked was founded in 2021 in San Francisco by Daniel Svonava, previously an ML tech lead for YouTube Ads at Google, and Ben Gutkovich. SIE grew out of the company's earlier work on search and embeddings. In the launch announcement, Svonava points to Serverless for jobs where someone else should hold the GPU quota, including evals, synthetic data generation, and public research, while self-hosting remains the route when data should stay in a customer's own infrastructure.