Hugging Face announced on June 26, 2026, that users can deploy private, OpenAI-compatible vLLM servers on its Jobs infrastructure using a single command. The system allows model deployment without provisioning servers or managing Kubernetes clusters, with billing calculated per second based on hardware usage.
Users must install huggingface_hub version 1.20.0 or higher, authenticate locally via hf auth login, and maintain an active payment method or credit balance before running a job. Executing the hf jobs run command deploys the official vllm/vllm-openai Docker image, using flags such as --flavor to allocate GPU hardware and --expose to route ports through Hugging Face's public jobs proxy.
Endpoints require a Hugging Face bearer token with read access to receive API requests, keeping the server gated to authorized account holders. Supported hardware options range from an a10g-large GPU instance priced at $1.50 per hour up to multi-GPU configurations like h200x2 and h200x8 for larger models such as Qwen3.5-122B.
Execution and Tool Integration
Developers can open an SSH shell directly into running job containers when passing the --ssh flag, provided their public SSH key is registered in their Hugging Face settings. Endpoints can also back terminal-based AI coding agents such as Pi when vLLM is launched with automatic tool choice enabled and configured with matching tool parsers.
Managed Inference Endpoints remain Hugging Face's alternative service for long-lived production workloads requiring scale-to-zero functionality, whereas HF Jobs targets flexible tasks like model evaluations and batch processing. The underlying port-exposure pattern on HF Jobs also supports alternative backends beyond vLLM, including llama.cpp and SGLang.
