# Connect NVIDIA NIM, hosted or your own

> This is the only cloud member where an address alone is a valid credential: an nvapi- key for the hosted endpoint, or a host for a server you run.

- Canonical: https://big-agi.com/docs/connect-nvidia-nim
- Minimum tier: open  [Open ⊂ Free ⊂ Pro - a tier only ever ADDS to the Open baseline]
- Kind: provider · Last content update generation: 2026-07-27 · Review by: 2026-10-27
- Answers: Where do I create an NVIDIA API key? · Can I point this at my own inference server instead? · Why is Direct Connection unavailable on the default host?

An address on its own counts as a credential here. Big-AGI accepts either an `nvapi-` key for NVIDIA's hosted endpoint or a host for a NIM or vLLM server you run, which may need no key at all. No other cloud AI service in this set can be configured without a key.

## Two ways to run it

The hosted endpoint is free with no payment method and rate-limited per account. Joining the developer programme is what yields the `nvapi-` key; confirm the current allowance at [build.nvidia.com](https://build.nvidia.com) rather than from any figure quoted elsewhere. A server you run yourself bills NVIDIA nothing, and its cost is hardware and electricity.

Free here means trial terms. NVIDIA may use prompts and outputs to improve its services, and production use is disallowed. The setup panel states the data-use half under the key field. Read the [trial terms](https://build.nvidia.com/terms) before putting work data through it.

## Getting an nvapi- key

1. Open [build.nvidia.com/settings/api-keys](https://build.nvidia.com/settings/api-keys), the page the app links, behind a sign-in.
2. Generate a key with the `Public API Endpoints` scope included, and pick its expiry date. It begins `nvapi-`.
3. Copy it. Running your own server instead? Skip the key and note its address.

> **Warning:**
> A key created without the `Public API Endpoints` scope passes every check Big-AGI makes, then fails all inference with `404 "Function not found for account"`. The expiry date you set is the other silent failure: a key that worked for months stops.

## A key, a host, or both

`Ctrl + Shift + M` opens **Models**; press **More Services** if the **Setup AI Models** wizard opens. Then **Add** -> **NVIDIA NIM**. For the hosted endpoint, paste into **NVIDIA API Key**. For your own server, fill **API Host** under **Advanced** instead - the pair validates on either one, so a keyless host is accepted. A NIM or vLLM box on your network goes in as its own address, `http://spark-xxxx.local:8000` for a DGX Spark. Press **Models** to fetch the list ([add and manage your keys](/docs/connect-models)).

Anything further from the standard shape belongs on [custom endpoints](/docs/connect-custom-endpoints).

## The curated model list

NVIDIA retires hosted models about every two weeks, and its raw model listing keeps returning the dead ids. Big-AGI drops the ones its own harvest confirmed retired, so they never reach your list and never fail mid-chat. A model the harvest has not seen yet carries a `[?]` prefix on its label. That prefix means uncurated: capabilities and context window stay unverified until the next harvest.

## The limit is a rate limit

On the hosted endpoint there is no bill to cap; the account's rate limit is what binds. The allowance counts requests, not tokens, so a long prompt costs the same as a short one. Big-AGI also spaces its own calls about 1.6 seconds apart there, to stay inside that limit, so multi-model runs feel slower on the hosted endpoint by design. A custom host is not throttled.

## Browser-direct only on your own host

Direct Connection here is inverted from the rest of the set: it can be turned on once you set a custom **API Host**. The default hosted endpoint admits only its own origins in the browser, so there the request goes through the Big-AGI fast edge servers.

> **Direct Connection works only with your API key stored in the browser**, and with an AI service that permits direct browser calls (CORS). Where it cannot be used, requests route through the Big-AGI fast edge servers instead - everything still works, within the standard upload size and time limits.

## When this doesn't work

| What you see | What it means | What to do |
|---|---|---|
| `Upstream responded with HTTP 403 Forbidden` | The hosted endpoint needs an `nvapi-` key and none was sent. | [Key and access errors](/docs/issue-key-and-access) |
| `404 Function not found for account` | The key lacks the `Public API Endpoints` scope, or that model is served only to entitled accounts. | Recreate the key with the scope; if it persists on one model only, use another: [key and access errors](/docs/issue-key-and-access) |

## Related

- [Connect an AI service](/docs/connect-models)
- [Connect any OpenAI-compatible endpoint](/docs/connect-custom-endpoints)
- [Direct Connection](/docs/feature-direct-connection)