HuggingFace Jobs adds one-command vLLM inference for any open-weight model
HuggingFace Jobs now spins up an OpenAI-compatible vLLM server for any supported open-weight model with a single CLI command, billed per-minute at $1.50/hour for an A10G and scaling to dual H200s for 100B+ models via tensor parallelism. No autoscaling is included — the fit is evals, batch generation, and self-hosted coding-agent backends; Inference Endpoints remains the production path.
Source: huggingface.co ↗
hf jobs run is docker run for HF infrastructure.
HuggingFace Jobs documentation
Why this matters
- → Spin up vLLM inference for any model with one command, no infrastructure setup.
- → Pay-per-minute pricing ($1.50/hr A10G) lowers cost for evals and batch jobs.
- → OpenAI-compatible API lets existing tools query models without code changes.
One-command model serving