Run a vLLM Server on HF Jobs in One Command

Hugging Face has introduced a method to run a private, OpenAI-compatible vLLM server on its infrastructure using a single command through HF Jobs. According to the source, the setup requires no server provisioning or Kubernetes, offering a pay-per-second model designed for quick testing, evaluations, and batch generation. Utilizing the official vllm/vllm-openai image, users specify a GPU flavor and expose the vllm port via a public jobs proxy, resulting in a gated endpoint that requires a Hugging Face token for authorization. The service scales to larger models by adjusting tensor parallel sizes and hardware flavors, such as using H200 chips for mixture-of-experts architectures. Developers can monitor and debug jobs interactively by opening an SSH shell inside the running container, provided their public key is registered. Furthermore, the endpoint can back terminal coding agents by enabling tool choices and tool-call parsers. Hugging Face distinguishes HF Jobs from its managed Inference Endpoints product, positioning Jobs as the flexible, control-heavy choice for experiments and one-off tasks, whereas Inference Endpoints provide production-grade features like finer-grained access control and scale-to-zero capabilities.