NVIDIA serves open models for free through an OpenAI-compatible API at https://integrate.api.nvidia.com/v1, authenticated with an API key from build.nvidia.com. Apus defaults the NVIDIA provider to Nemotron 3 Ultra, NVIDIA's 550B total / 55B active reasoning model for long-context agentic work.

Getting started

Create an API key at [build.nvidia.com](https://build.nvidia.com/settings/api-keys).

```bash
export NVIDIA_API_KEY="nvapi-..."
apus onboard --auth-choice nvidia-api-key
```

```bash
apus models set nvidia/nvidia/nemotron-3-ultra-550b-a55b
```

For non-interactive setup, pass the key directly:

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

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

When an NVIDIA API key is configured, setup and model-selection paths check https://integrate.api.nvidia.com/v1/models for available model IDs, cached for 30 seconds. NVIDIA's public https://assets.ngc.nvidia.com/products/api-catalog/featured-models.json feed provides ranking and token limits, cached for 24 hours. Featured models appear first only while the inference inventory still lists them; other available bundled chat models follow. A fresh inventory can restore a previously hidden model that NVIDIA has republished.

The hosted catalog applies the same feed: NVIDIA's featured models lead its recommended models, followed by the global list.

The inventory also contains embeddings and other non-chat endpoints, without capability metadata. Apus therefore offers only exact models with bundled chat metadata or valid featured-model metadata; it does not guess capabilities from model names. Unknown IDs can still be configured explicitly; listing alone does not prove chat compatibility. This is not a complete automatic catalog of every NVIDIA model.

Both public fetches use fixed HTTPS hosts and send no credentials. A failed inventory or featured request marks discovery unavailable and retains the last successful catalog for the same provider configuration and credentials. Failed featured metadata cannot silently remove previously discovered models. A successful empty inventory clears discovered models, even if the featured feed fails. Without NVIDIA auth, browsing uses the bundled catalog without fetching.

Nemotron 3.5 Lightning

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

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

Lightning is selectable when the live inventory lists it even if it is absent from the featured feed. Nemotron 3 Ultra remains the default.

Nemotron 3 Ultra

Nemotron 3 Ultra is the default NVIDIA model in Apus. NVIDIA's build page for nvidia/nemotron-3-ultra-550b-a55b lists it as an available free endpoint with a 1M-token context specification.

The bundled Ultra row sends chat_template_kwargs: { enable_thinking: false, force_nonempty_content: true } by default so normal chat output stays in the visible answer instead of exposing reasoning text.

Use Ultra for the highest-capability NVIDIA default. Select Nemotron 3.5 Lightning or Nemotron 3 Super when you want a smaller Nemotron option, or choose one of the third-party models hosted in NVIDIA's catalog when their context, latency, or behavior fits better.

Bundled fallback catalog

The bundled rows provide known chat metadata and an offline fallback. Deprecated compatibility rows keep existing exact model references recognizable but stay out of model pickers.

The full compatibility catalog also retains these shipped refs for existing configurations 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 hidden from bundled and offline model pickers unless NVIDIA republishes them in its inference inventory. NVIDIA has retired the Qwen endpoint, so requests using its model reference no longer work. Migrate existing Qwen configurations to an active model.

Advanced configuration

The provider auto-enables when the `NVIDIA_API_KEY` environment variable is
set or a key was stored during onboarding. No explicit provider config is
required beyond the key.

Apus uses NVIDIA's inference inventory for availability and its featured
feed for ranking. Exact bundled metadata preserves reasoning and image
capabilities omitted by the featured feed. Deprecated exact-reference
compatibility rows stay hidden from the offline fallback; fresh inventory
can restore models that NVIDIA has republished. Costs default to `0` in source
since NVIDIA currently offers free API access for the listed models.

Apus talks to NVIDIA with the `openai-completions` adapter against the
standard `/v1` chat completions route. Any OpenAI-compatible tooling should
work out of the box with the NVIDIA base URL.

NVIDIA's Ultra sample request uses `chat_template_kwargs.enable_thinking`
and `reasoning_budget` for reasoning output. Apus's bundled Ultra row
disables template thinking by default for normal chat use. If you need to
opt into NVIDIA reasoning output or force other NVIDIA-specific request
fields, set per-model params and keep provider-specific overrides scoped to
the NVIDIA model:

```json5
{
  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` merges into any `chat_template_kwargs`
already on the request instead of replacing the whole object.
`params.extra_body` is the final OpenAI-compatible request-body override
and overwrites colliding payload keys, so use it only for fields NVIDIA
documents for the selected endpoint.

Some NVIDIA-hosted custom models can take longer than the default ~120s
model idle watchdog before they emit a first response chunk. For custom
NVIDIA provider entries, raise the provider timeout instead of the whole
agent runtime timeout; `timeoutSeconds` covers provider HTTP requests and
raises the idle/stream watchdog ceiling for that provider. The provider id
below (`custom-integrate-api-nvidia-com`) is a name you choose, not a
reserved value; any id works as long as your model refs use the same prefix:

```json5
{
  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" },
        },
      },
    },
  },
}
```

Related

Choosing providers, model refs, and failover behavior.

Full config reference for agents, models, and providers.