NVIDIA Provider Setup and Configuration in OpenClaw

NVIDIA offers free access to open models through an API that follows OpenAI's interface, located at https://integrate.api.nvidia.com/v1. Authentication relies on an API key obtained from build.nvidia.com. By default, OpenClaw assigns Nemotron 3 Ultra as the NVIDIA provider, a reasoning model with 550B total and 55B active parameters designed for extended-context agent tasks.

Getting started

Get your API key

Generate your 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

For environments where interactive setup isn't an option, supply the key directly:

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

Warning

Using --nvidia-api-key leaves the key visible in shell history and ps output. Whenever possible, go with 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 is in place, the setup and model-selection flows query https://integrate.api.nvidia.com/v1/models to discover available model IDs, with results held in cache for 30 seconds. NVIDIA's public https://assets.ngc.nvidia.com/products/api-catalog/featured-models.json feed supplies ranking data and token limits, cached for 24 hours. Featured models take priority in the listing only while the inference inventory still includes them; other bundled chat models that are available come after. A refreshed inventory can bring back a model that NVIDIA has republished after it was previously hidden.

Embeddings and other non-chat endpoints appear in the inventory as well, but they carry no capability metadata. Because of this, OpenClaw only surfaces exact models that have bundled chat metadata or valid featured-model metadata; it never infers capabilities from model names. Explicit configuration of unknown IDs remains possible, and appearing in the list alone doesn't guarantee chat support. This isn't a full automatic catalog covering every NVIDIA model.

Both public requests go to fixed HTTPS hosts without any credentials attached. When the featured feed fails, bundled chat models that are available still show up. If the inventory is missing or malformed, browsing and setup fall back to the bundled option, while the fresh-live catalog remains empty. A successful but empty inventory won't bring back outdated bundled entries. Without NVIDIA authentication, browsing relies on the bundled catalog and skips fetching.

Nemotron 3.5 Lightning

nvidia/nemotron-3.5-lightning-30b-a3b is NVIDIA's reasoning model for agent work, smaller at 30B total and 3B active parameters. The bundled entry notes its 1M context and a 16,384-token output cap that matches NVIDIA's hosted example. To pick it:

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

Lightning becomes available when the live inventory lists it, even if the featured feed doesn't include it. Nemotron 3 Ultra stays as the default choice.

Nemotron 3 Ultra

In OpenClaw, Nemotron 3 Ultra serves as the default NVIDIA model. NVIDIA's build page for nvidia/nemotron-3-ultra-550b-a55b shows it as a free endpoint with a 1M-token context.

The bundled Ultra row sends chat_template_kwargs: { enable_thinking: false, force_nonempty_content: true } by default, keeping standard chat output in the visible answer rather than exposing reasoning text.

Choose Ultra when you want the strongest NVIDIA default. Stick with Super if the smaller Nemotron 3 option fits, or go with a third-party model from NVIDIA's catalog when its context, latency, or behavior suits your needs better.

Bundled fallback catalog

Known chat metadata and an offline fallback come from the bundled rows. Deprecated compatibility rows keep existing exact model references working for recognition but remain absent from model pickers.

The full compatibility catalog also keeps these shipped refs for existing setups and migration: nvidia/qwen/qwen3.5-397b-a17b, nvidia/moonshotai/kimi-k2.5, nvidia/z-ai/glm-5.1, nvidia/z-ai/glm5, and nvidia/minimaxai/minimax-m2.7. These references stay out of bundled and offline model pickers unless NVIDIA republishes them in its inference inventory. Since NVIDIA has retired the Qwen endpoint, requests using that model reference no longer work. Move existing Qwen configurations to an active model.

Advanced configuration

Auto-enable behavior

Setting the NVIDIA_API_KEY environment variable or storing a key during onboarding auto-enables this provider. No extra provider configuration is needed beyond the key itself.

Catalog and pricing

Availability comes from NVIDIA's inference inventory, while ranking relies on the featured feed. Exact bundled metadata keeps reasoning and image capabilities that the featured feed omits. Deprecated exact-reference compatibility rows remain hidden from the offline fallback; a fresh inventory can restore models NVIDIA has republished. Costs default to 0 in source since NVIDIA currently provides free API access for these models.

OpenAI-compatible endpoint

OpenClaw connects to NVIDIA using the openai-completions adapter against the standard /v1 chat completions endpoint. Any tooling that works with OpenAI should function without changes when pointed at the NVIDIA base URL.

Nemotron 3 Ultra reasoning params

NVIDIA's Ultra sample request employs chat_template_kwargs.enable_thinking and reasoning_budget for reasoning output. OpenClaw's bundled Ultra row turns off template thinking by default for regular chat. To opt into NVIDIA reasoning output or push 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 },
          },
        },
      },
    },
  },
}

params.chat_template_kwargs combines with any chat_template_kwargs already on the request rather than replacing the entire object. params.extra_body serves as the final OpenAI-compatible request-body override and overwrites conflicting payload keys, so reserve it for fields NVIDIA documents for the chosen endpoint.

Slow custom provider responses

Certain NVIDIA-hosted custom models may take longer than the default ~120s model idle watchdog before sending the first response chunk. For custom NVIDIA provider entries, increase the provider timeout rather than the whole agent runtime timeout; timeoutSeconds covers provider HTTP requests and lifts the idle/stream watchdog ceiling for that provider:

{
  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

NVIDIA models are free to use at the moment. 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 covering agent, model, and provider settings.