Inference lets you run a model, an app, or a workflow as a long-running service on a node operator's GPU, reachable over HTTP for as long as you've paid for it. It's the difference between a one-off compute job and something that stays live and answers requests.

No. Curated Models and Templates are point-and-click: pick a package or an outcome, choose an environment, and pay. Coding knowledge only matters if you use Custom Models to set launch flags yourself, or if you're calling the API from your own app.

No. You connect a wallet with funds on Base and pay node operators directly. There's no signup, no instance provisioning, and no per-seat licence.

Every entry point walks the same four stages, though some are skipped when they don't apply: pick what to run, choose an environment that can handle it, configure anything the workload needs (launch flags, an access token, a storage bucket), then pay and launch. Curated packages, Services, and Templates size the hardware for you and can skip straight to payment; Custom Models always walks the full sequence since there's nothing pre-declared to fall back on.

A wallet with funds on Base (an external wallet or the built-in Smart Wallet via email or social login), and a rough sense of what your model needs. For Curated Models, the catalogue filters packages against your required resources before you pick one, Custom Models and Services let you configure anything, so that check doesn't apply there. A Hugging Face token is only needed for gated or private models.

Curated Models are pre-sized packages, pick one and go straight to payment. Custom Models let you configure every launch flag yourself on any Hugging Face model. Services are empty apps (ComfyUI, Open WebUI, JupyterLab) you load your own models into. Templates are Services with models and workflows already installed, chosen by outcome rather than assembled by hand.