One Command to a vLLM Endpoint: HuggingFace Jobs; Copilot Benchmarks Its Harness Across Models

TL;DR: HuggingFace Jobs now spins up a private OpenAI-compatible vLLM endpoint in one shell command; GitHub's benchmark shows its agentic harness matches vendor harnesses while using fewer tokens.

HuggingFace Jobs + vLLM. A new HuggingFace blog post (June 26) shows how to launch a pay-per-second OpenAI-compatible inference endpoint with a single command:

hf jobs run --flavor a10g-large --expose 8000 --timeout 2h \
  vllm/vllm-openai:latest \
  vllm serve Qwen/Qwen3-4B --host 0.0.0.0 --port 8000

An A10G instance runs at $1.50/hr with per-second billing — no servers to provision, no Kubernetes. The endpoint is OpenAI-compatible, so existing client code works unchanged. Run hf jobs hardware to see available flavors and pricing.

GitHub Copilot agentic harness benchmark. GitHub published benchmarks (June 26) of its shared orchestration layer — the harness that powers Copilot CLI, code review, and IDE features — against Claude Sonnet 4.6, Opus 4.7, GPT-5.4, and GPT-5.5 on TerminalBench 2.0. Result: Copilot's harness matched or beat model-vendor-native harnesses on task completion while consuming fewer tokens.

Why it matters: HF Jobs removes the last friction point for self-hosting — a private vLLM endpoint is now a one-liner at ~$1.50/hr, no infra knowledge needed. The Copilot harness data is worth internalizing: token-efficient orchestration (tool selection, context management) matters as much as model choice on real agentic tasks.

Sources: Run a vLLM Server on HF Jobs in One Command — HuggingFace Blog | Evaluating GitHub Copilot Agentic Harness Across Models — GitHub Blog