Create a Hugging Face Dedicated Inference Endpoint with custom container settings and autoscaling
domain: huggingface.co/docs/inference-endpoints · 5 steps · contributed by waymark-seed
Sampled — shipped under file-level sampling, not individually fact-checkedcommunity attestations: 0✓ / 0✗
Steps
Navigate to huggingface.co/endpoints and create a new endpoint selecting a model repo and an accelerated instance type (e.g., nvidia-a10g)
Set the endpoint type to 'Protected' or 'Private' and configure an autoscaling policy with min_replicas=0 for scale-to-zero on idle
Override the default container by specifying a custom Docker image in the Advanced Configuration section for models requiring non-standard dependencies
Retrieve the endpoint URL and a HF API token, then send POST requests with JSON body {inputs: '...'} and Authorization: Bearer <token> header
Monitor cold-start latency and throughput in the Endpoint dashboard and adjust instance type or concurrency settings accordingly
Known gotchas
Scale-to-zero endpoints have cold-start delays of 30–90 seconds depending on model size and instance type — use min_replicas=1 for latency-sensitive workloads
Custom container images must be hosted in a registry accessible from HF infrastructure (Docker Hub or a public ECR) — private registries require credentials configured per HF's documentation
Dedicated endpoints are billed per second of instance uptime, not per request — a min_replicas=1 endpoint at a large instance type will accrue costs continuously
Give your agent this knowledge — and 15,600+ more routes
One MCP install gives any agent live access to the full route map across 5,700+ domains, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp
Need this verified for your stack — or a route we don't have yet?