OpenClaw and NVIDIA

NVIDIA serves open models for free through an OpenAI-compatible API, and OpenClaw has a provider page for it: an API key from NVIDIA’s build portal, one onboarding choice, and a default model, Nemotron 3 Ultra, a reasoning model with 550 billion total parameters and 55 billion active, meant for long-context agentic work. That is the whole setup. The rest of this post is what free means here, and the two cautions the docs and the security pages add.

Setup

• Create an API key on the NVIDIA build portal.
• Export it as the NVIDIA key variable and run onboarding with the NVIDIA auth choice; a flag passes the key directly for non-interactive setup.
• Set the model with the models command; the default is Nemotron 3 Ultra.
• The config example is an environment variable plus a provider block; the endpoint is NVIDIA’s OpenAI-compatible base URL.

The nvidia-api-key flag lands the key in shell history and ps output. Prefer the NVIDIA_API_KEY environment variable when possible.

What free means

Free here is hosted, not local: NVIDIA runs the model and serves it through its API, so there is no hardware to buy and no memory floor on your machine, unlike the local route Free and local models for OpenClaw describes. The docs list it under open models for free, and a free hosted tier is exactly the kind of thing that changes without an OpenClaw release, so treat the price as today’s.

The two cautions

• The tier rule: an open model that is free is still a model tier, and the prompt-injection page says to use the best current tier for anything that runs tools or reads untrusted content. Nemotron is positioned for agentic work; measure it on your own tasks before it runs your inbox.
• The key: a key in shell history is a key in every backup of that history; the environment variable is the documented preference, and a vault is better than either.

Prompt injection and OpenClaw is the page behind the first caution, and Which model for OpenClaw is the general guide to choosing between the providers.

On Diali

Diali’s managed models are the fifteen included in every plan, spent from a credit balance, and bring-your-own-key covers ten providers from the vault; the dashboard lists both. Hosted OpenClaw on Diali is the assistant; the AI credit calculator turns expected use into a number.

• One key, one onboarding choice, Nemotron by default.
• Free and hosted, not free and local.
• A free tier is still a tier: measure before it runs tools.