NVIDIA Provider Setup for OpenClaw: API Key and Models

NVIDIA offers open models at no cost via an OpenAI-compatible API hosted at https://integrate.api.nvidia.com/v1. Authentication uses an API key obtained from build.nvidia.com. For the NVIDIA provider, OpenClaw ships with Nemotron 3 Ultra as the default: NVIDIA's reasoning model with 550B total and 55B active parameters, built for long-context agentic workloads.

Getting started

Get your API key

Generate an API key on build.nvidia.com.

Export the key and run onboarding

export NVIDIA_API_KEY="nvapi-..."
openclaw onboard --auth-choice nvidia-api-key

Set an NVIDIA model

openclaw models set nvidia/nvidia/nemotron-3-ultra-550b-a55b

To configure without prompts, supply the key on the command line:

openclaw onboard --auth-choice nvidia-api-key --nvidia-api-key "nvapi-..."

Warning

The key ends up recorded in shell history by --nvidia-api-key and in the output of ps. Wherever you can, use the NVIDIA_API_KEY environment variable instead.

Config example

{
  env: { vars: { NVIDIA_API_KEY: "nvapi-..." } },
  models: {
    providers: {
      nvidia: {
        baseUrl: "https://integrate.api.nvidia.com/v1",
        api: "openai-completions",
      },
    },
  },
  agents: {
    defaults: {
      model: { primary: "nvidia/nvidia/nemotron-3-ultra-550b-a55b" },
    },
  },
}

Live model catalog

Once an NVIDIA API key exists, the setup and model-selection code queries https://integrate.api.nvidia.com/v1/models for the model IDs on offer, with a 30-second cache. Ranking data and token limits come from NVIDIA's public https://assets.ngc.nvidia.com/products/api-catalog/featured-models.json feed, cached for 24 hours. Featured models are listed first, but only as long as the inference inventory still contains them; the remaining bundled chat models that are available come after. If NVIDIA republishes a model that was previously hidden, a refreshed inventory can bring it back.

The hosted catalog draws on the same feed: NVIDIA's featured models come first on its recommended models page, with the global list after them.

Embeddings and other non-chat endpoints also appear in the inventory, and they carry no capability metadata. Because of that, OpenClaw lists only exact models that have bundled chat metadata or valid featured-model metadata. It never infers capabilities from a model's name. You can still configure unknown IDs by hand; being listed does not by itself prove a model supports chat. The result is not a full automatic catalog of every NVIDIA model.

Both public fetches go to fixed HTTPS hosts and carry no credentials. If an inventory or featured request fails, discovery is marked unavailable and the last catalog that succeeded is kept for the same provider configuration and credentials. A failed featured metadata fetch cannot quietly drop models discovered earlier. A successful but empty inventory does clear discovered models, even when the featured feed fails. When no NVIDIA auth is present, browsing relies on the bundled catalog and performs no fetching.

Nemotron 3.5 Lightning

nvidia/nemotron-3.5-lightning-30b-a3b is NVIDIA's smaller reasoning model for agentic work, with 30B total and 3B active parameters. Its bundled entry records a 1M context and a 16,384-token output budget, matching NVIDIA's hosted example. To select it:

openclaw models set nvidia/nvidia/nemotron-3.5-lightning-30b-a3b

Lightning can be chosen whenever the live inventory lists it, even if the featured feed omits it. Nemotron 3 Ultra stays the default.

Nemotron 3 Ultra

Within OpenClaw, Nemotron 3 Ultra is the default NVIDIA model. On NVIDIA's build page, nvidia/nemotron-3-ultra-550b-a55b appears as a free endpoint that is available, with a 1M-token context specification.

By default the bundled Ultra entry sends chat_template_kwargs: { enable_thinking: false, force_nonempty_content: true }, which keeps ordinary chat output in the visible answer rather than exposing reasoning text.

Pick Ultra when you want the most capable NVIDIA default. For a smaller text-only Nemotron, choose Nemotron 3.5 Lightning; for image and video understanding, choose Nemotron 3.5 Super VL; or pick any third-party model hosted in NVIDIA's catalog if its context, latency, or behavior suits you better.

Nemotron 3.5 Super VL

Nemotron 3.5 Super VL is NVIDIA's vision-language model, accepting text, image, and video input.

To use it for text, image, and video input:

openclaw models set nvidia/nvidia/nemotron-3.5-super-vl-120b-a12b

Combine Super VL's image handling with Nemotron 3 Ultra for text: make Ultra the primary model and Super VL the image model.

{
  "agents": {
    "defaults": {
      "model": { "primary": "nvidia/nvidia/nemotron-3-ultra-550b-a55b" },
      "imageModel": { "primary": "nvidia/nvidia/nemotron-3.5-super-vl-120b-a12b" }
    }
  }
}

When the primary model cannot accept images, OpenClaw falls back to imageModel for image understanding.

Bundled fallback catalog

The bundled entries supply known chat metadata plus an offline fallback. Retired models are left out, and existing configurations are never rewritten automatically.

Should a configured model have been retired, choose an active one from openclaw models list --provider nvidia.

Advanced configuration

Auto-enable behavior

The provider turns on by itself when the NVIDIA_API_KEY environment variable is present or a key was saved during onboarding. Apart from the key, no explicit provider config is needed.

Catalog and pricing

Availability comes from NVIDIA's inference inventory, and ranking comes from its featured feed. Exact bundled metadata keeps the reasoning and image capabilities that the featured feed leaves out. Retired models are dropped from the offline fallback, while fresh featured metadata and inventory can restore any model NVIDIA has republished. Costs are set to 0 in source by default, because NVIDIA presently provides free API access for the listed models.

OpenAI-compatible endpoint

OpenClaw reaches NVIDIA through the openai-completions adapter, calling the standard /v1 chat completions route. Any OpenAI-compatible tooling ought to work as-is against the NVIDIA base URL.

Nemotron 3 Ultra reasoning params

NVIDIA's Ultra sample request relies on chat_template_kwargs.enable_thinking and reasoning_budget to produce reasoning output. For normal chat use, OpenClaw's bundled Ultra entry turns template thinking off by default. To opt into NVIDIA reasoning output, or to force other NVIDIA-specific request fields, set per-model params and keep provider-specific overrides limited to the NVIDIA model:

{
  agents: {
    defaults: {
      models: {
        "nvidia/nvidia/nemotron-3-ultra-550b-a55b": {
          params: {
            chat_template_kwargs: { enable_thinking: true },
            extra_body: { reasoning_budget: 16384 },
          },
        },
      },
    },
  },
}

Rather than replacing the whole object, params.chat_template_kwargs merges into any chat_template_kwargs already present on the request. As the last OpenAI-compatible override of the request body, params.extra_body overwrites colliding payload keys, so restrict it to fields that NVIDIA documents for the chosen endpoint.

Slow custom provider responses

Certain NVIDIA-hosted custom models may need longer than the default ~120s model idle watchdog before sending their first response chunk. For custom NVIDIA provider entries, increase the provider timeout rather than the timeout for the whole agent runtime; timeoutSeconds governs provider HTTP requests and lifts the idle/stream watchdog ceiling for that provider. The provider id shown below (custom-integrate-api-nvidia-com) is a name of your choosing, not a reserved value; any id will do, provided your model refs share the same prefix:

{
  models: {
    providers: {
      "custom-integrate-api-nvidia-com": {
        baseUrl: "https://integrate.api.nvidia.com/v1",
        api: "openai-completions",
        apiKey: "NVIDIA_API_KEY",
        timeoutSeconds: 300,
      },
    },
  },
  agents: {
    defaults: {
      models: {
        "custom-integrate-api-nvidia-com/meta/llama-3.1-70b-instruct": {
          params: { thinking: "off" },
        },
      },
    },
  },
}

Tip

At present, NVIDIA models cost nothing to use. See build.nvidia.com for current availability and rate-limit details.

Related

• Model selection, Picking providers, model refs, and failover behavior.

• Configuration reference: Complete reference for agent, model, and provider settings.