> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/NVIDIA/NemoClaw/llms.txt
> Use this file to discover all available pages before exploring further.

# Inference profiles

> Configuration reference for NemoClaw inference providers — NVIDIA cloud, NCP partners, and experimental local options.

# Inference profiles

NemoClaw ships with inference profiles defined in `blueprint.yaml`. Each profile configures an OpenShell inference provider and model route. The agent inside the sandbox uses whichever provider and model is active. Inference requests are routed transparently through the OpenShell gateway — they never leave the sandbox directly.

<Note>
  **Two setup paths, two provider names.** When using `openclaw nemoclaw launch` or `openclaw nemoclaw migrate`, the blueprint creates providers with the names in the table below (for example, `nvidia-inference`). When using the standalone `nemoclaw onboard` wizard, the provider is named `nvidia-nim` instead. The provider name matters when running `openshell inference set` to switch models.
</Note>

## Profile summary

| Profile     | Provider name      | Model                               | Endpoint                   | Notes                                          |
| ----------- | ------------------ | ----------------------------------- | -------------------------- | ---------------------------------------------- |
| `default`   | `nvidia-inference` | `nvidia/nemotron-3-super-120b-a12b` | `integrate.api.nvidia.com` | Production. Requires NVIDIA API key.           |
| `ncp`       | `nvidia-ncp`       | `nvidia/nemotron-3-super-120b-a12b` | Configurable               | NCP partner endpoint. Requires NVIDIA API key. |
| `nim-local` | `nim-local`        | `nvidia/nemotron-3-super-120b-a12b` | `nim-service.local:8000`   | Experimental. Requires NIM API key.            |
| `vllm`      | `vllm-local`       | `nvidia/nemotron-3-nano-30b-a3b`    | `localhost:8000`           | Experimental. No API key required.             |

## Available models

The `nvidia-inference` provider registers the following models from [build.nvidia.com](https://build.nvidia.com):

| Model ID                                   | Label                   | Context window | Max output |
| ------------------------------------------ | ----------------------- | -------------- | ---------- |
| `nvidia/nemotron-3-super-120b-a12b`        | Nemotron 3 Super 120B   | 131,072        | 8,192      |
| `nvidia/llama-3.1-nemotron-ultra-253b-v1`  | Nemotron Ultra 253B     | 131,072        | 4,096      |
| `nvidia/llama-3.3-nemotron-super-49b-v1.5` | Nemotron Super 49B v1.5 | 131,072        | 4,096      |
| `nvidia/nemotron-3-nano-30b-a3b`           | Nemotron 3 Nano 30B     | 131,072        | 4,096      |

The `default` profile activates Nemotron 3 Super 120B. You can switch to any model in the catalog at runtime without restarting the sandbox.

## Provider types

<Tabs>
  <Tab title="NVIDIA Build (default)">
    The default profile routes inference to NVIDIA's hosted API at [build.nvidia.com](https://build.nvidia.com). This is the recommended option for most users — it requires no local infrastructure.

    **Profile:** `default`

    | Field          | Value                                 |
    | -------------- | ------------------------------------- |
    | Provider name  | `nvidia-inference`                    |
    | Provider type  | `nvidia`                              |
    | Endpoint       | `https://integrate.api.nvidia.com/v1` |
    | Default model  | `nvidia/nemotron-3-super-120b-a12b`   |
    | Credential env | `NVIDIA_API_KEY`                      |

    **Getting an API key:**

    1. Go to [build.nvidia.com](https://build.nvidia.com) and sign in.
    2. Navigate to **Settings → API Keys**.
    3. Create a new key and copy it.
    4. Pass it to `openclaw nemoclaw onboard --api-key <key>` or set `NVIDIA_API_KEY` in your environment.

    The onboard wizard stores the key in `~/.nemoclaw/credentials.json` (mode `600`).

    **Onboard:**

    ```bash theme={null}
    openclaw nemoclaw onboard --endpoint build --api-key nvapi-xxxx --model nvidia/nemotron-3-super-120b-a12b
    ```

    **Switch model at runtime:**

    ```bash theme={null}
    openshell inference set --provider nvidia-inference --model nvidia/llama-3.3-nemotron-super-49b-v1.5
    ```

    The change takes effect immediately. No sandbox restart is needed.
  </Tab>

  <Tab title="NCP partner">
    The NCP (NVIDIA Cloud Partner) profile routes inference to a dedicated partner endpoint with configurable capacity and SLA backing.

    **Profile:** `ncp`

    | Field            | Value                               |
    | ---------------- | ----------------------------------- |
    | Provider name    | `nvidia-ncp`                        |
    | Provider type    | `nvidia`                            |
    | Endpoint         | Configurable (partner-supplied URL) |
    | Default model    | `nvidia/nemotron-3-super-120b-a12b` |
    | Credential env   | `NVIDIA_API_KEY`                    |
    | Dynamic endpoint | `true`                              |

    **Onboard:**

    ```bash theme={null}
    openclaw nemoclaw onboard \
      --endpoint ncp \
      --ncp-partner acme \
      --endpoint-url https://acme.api.nvidia.com/v1 \
      --api-key nvapi-xxxx \
      --model nvidia/nemotron-3-super-120b-a12b
    ```

    The `--ncp-partner` flag records the partner name for reference. The `--endpoint-url` sets the dynamic endpoint for the NCP provider.
  </Tab>

  <Tab title="Local NIM (experimental)">
    <Warning>
      `nim-local` is experimental and may change without notice. Set `NEMOCLAW_EXPERIMENTAL=1` to enable it in the interactive onboard menu.
    </Warning>

    Routes inference to a self-hosted NIM (NVIDIA Inference Microservice) container running on the local network.

    **Profile:** `nim-local`

    | Field            | Value                               |
    | ---------------- | ----------------------------------- |
    | Provider name    | `nim-local`                         |
    | Provider type    | `openai`                            |
    | Default endpoint | `http://nim-service.local:8000/v1`  |
    | Default model    | `nvidia/nemotron-3-super-120b-a12b` |
    | Credential env   | `NIM_API_KEY`                       |

    **Onboard:**

    ```bash theme={null}
    NEMOCLAW_EXPERIMENTAL=1 openclaw nemoclaw onboard \
      --endpoint nim-local \
      --endpoint-url http://nim-service.local:8000/v1 \
      --api-key your-nim-key \
      --model nvidia/nemotron-3-super-120b-a12b
    ```

    Credential validation is best-effort during onboarding — the NIM service may not be running yet.
  </Tab>

  <Tab title="vLLM (experimental)">
    <Warning>
      `vllm` is experimental and intended for local development. Set `NEMOCLAW_EXPERIMENTAL=1` to enable it.
    </Warning>

    Routes inference to a local vLLM server running on the host.

    **Profile:** `vllm`

    | Field              | Value                                         |
    | ------------------ | --------------------------------------------- |
    | Provider name      | `vllm-local`                                  |
    | Provider type      | `openai`                                      |
    | Endpoint           | `http://localhost:8000/v1` (via host gateway) |
    | Default model      | `nvidia/nemotron-3-nano-30b-a3b`              |
    | Credential env     | `OPENAI_API_KEY`                              |
    | Default credential | `dummy`                                       |

    The endpoint is resolved through `http://host.openshell.internal:8000/v1` so sandbox network policy is not required for local traffic.

    **Onboard:**

    ```bash theme={null}
    NEMOCLAW_EXPERIMENTAL=1 openclaw nemoclaw onboard \
      --endpoint vllm \
      --model nvidia/nemotron-3-nano-30b-a3b
    ```
  </Tab>

  <Tab title="Ollama (experimental)">
    <Warning>
      `ollama` is experimental. Set `NEMOCLAW_EXPERIMENTAL=1` to enable it.
    </Warning>

    Routes inference to a local Ollama server on the host.

    **Profile:** `ollama` (experimental, not listed in `blueprint.yaml`; configured directly via `openclaw nemoclaw onboard`)

    | Field              | Value                                     |
    | ------------------ | ----------------------------------------- |
    | Provider name      | `ollama-local`                            |
    | Provider type      | `openai`                                  |
    | Default endpoint   | `http://host.openshell.internal:11434/v1` |
    | Credential env     | `OPENAI_API_KEY`                          |
    | Default credential | `ollama`                                  |

    If Ollama is detected running on `localhost:11434` and `NEMOCLAW_EXPERIMENTAL=1` is set, the onboard wizard selects it automatically.

    **Onboard:**

    ```bash theme={null}
    NEMOCLAW_EXPERIMENTAL=1 openclaw nemoclaw onboard --endpoint ollama
    ```
  </Tab>

  <Tab title="Custom (experimental)">
    <Warning>
      `custom` is experimental. Set `NEMOCLAW_EXPERIMENTAL=1` to enable it.
    </Warning>

    Routes inference to any OpenAI-compatible endpoint.

    | Field          | Value                           |
    | -------------- | ------------------------------- |
    | Provider name  | `nvidia-ncp`                    |
    | Endpoint       | Configurable (`--endpoint-url`) |
    | Credential env | `NVIDIA_API_KEY`                |

    **Onboard:**

    ```bash theme={null}
    NEMOCLAW_EXPERIMENTAL=1 openclaw nemoclaw onboard \
      --endpoint custom \
      --endpoint-url https://my-llm-server.internal/v1 \
      --api-key my-key \
      --model my-custom-model
    ```
  </Tab>
</Tabs>

## Enabling experimental providers

Local inference options (NIM, vLLM, Ollama, custom) are hidden by default. Set the `NEMOCLAW_EXPERIMENTAL` environment variable to expose them in the interactive menu:

```bash theme={null}
NEMOCLAW_EXPERIMENTAL=1 openclaw nemoclaw onboard
```

You can also pass `--endpoint` directly without the environment variable — NemoClaw will issue a warning but proceed:

```bash theme={null}
openclaw nemoclaw onboard --endpoint nim-local --endpoint-url http://nim-service.local:8000/v1 --api-key key
```

## Credentials

Credentials are stored in `~/.nemoclaw/credentials.json` with file permissions set to `600`. The credential environment variable used depends on the endpoint type:

| Endpoint type | Credential env                       |
| ------------- | ------------------------------------ |
| `build`       | `NVIDIA_API_KEY`                     |
| `ncp`         | `NVIDIA_API_KEY`                     |
| `custom`      | `NVIDIA_API_KEY`                     |
| `nim-local`   | `NIM_API_KEY`                        |
| `vllm`        | `OPENAI_API_KEY` (default: `dummy`)  |
| `ollama`      | `OPENAI_API_KEY` (default: `ollama`) |

## Switching models at runtime

After the sandbox is running, switch the active model with the OpenShell CLI. The provider name depends on the setup method used:

<Tabs>
  <Tab title="Plugin setup (openclaw nemoclaw)">
    The blueprint creates a provider named `nvidia-inference`:

    ```bash theme={null}
    openshell inference set --provider nvidia-inference --model nvidia/llama-3.1-nemotron-ultra-253b-v1
    ```
  </Tab>

  <Tab title="Standalone setup (nemoclaw onboard)">
    The standalone wizard creates a provider named `nvidia-nim`:

    ```bash theme={null}
    openshell inference set --provider nvidia-nim --model nvidia/llama-3.1-nemotron-ultra-253b-v1
    ```
  </Tab>
</Tabs>

The change takes effect immediately. No sandbox restart is needed.