NVIDIA NIM API endpoint URL, and how to call it for free

POST https://integrate.api.nvidia.com/v1/chat/completions

POST https://dreamprompting.com/api/v1/chat/completions

If you want a free NVIDIA NIM API without a payment method or a management console, you can reach NIM models directly through the DreamPrompting gateway. You need one free API key, which takes a few seconds to get by signing in with Google, with no credit card. It is an OpenAI-compatible endpoint that routes your request to NVIDIA NIM (and other providers) with automatic failover, so you can call Llama, Nemotron, and more from one base URL. This post explains what NIM is, which models you can reach, how to call them for free, and the one speed tradeoff to plan around.

What is NVIDIA NIM?

NVIDIA NIM (NVIDIA Inference Microservices) is a set of prebuilt, GPU-accelerated inference containers that serve popular open models behind a standard API. Instead of building your own serving stack, NIM packages the model, the runtime, and the optimizations together so it can run efficiently on NVIDIA hardware. NIM exposes a large catalog of models, and a free NVIDIA NIM API through DreamPrompting lets you tap that catalog without touching any infrastructure yourself.

Which NIM models can you reach?

NVIDIA NIM is a live provider on the gateway, and it offers many models. Common families you can reach include the following:

• Meta Llama chat and instruct models (for example a small, fast llama-3.1-8b-instruct).
• NVIDIA Nemotron models tuned for reasoning and instruction following.
• Additional open models that NVIDIA publishes as NIM microservices over time.

Because the catalog changes, the model ids you can pin are best read live. Ask the gateway what is available with a single call to the models list, shown below. This is the reliable way to confirm the exact NIM id you want before you hardcode it, and it saves you from guessing at names that may have been renamed or retired.

curl https://dreamprompting.com/api/v1/models \
  -H "Authorization: Bearer YOUR_API_KEY"

How to use the free NVIDIA NIM API

The gateway follows the OpenAI chat format, so any tool that speaks that format works. Create your free API key on your account page, send it as Authorization: Bearer YOUR_API_KEY, point the base URL at https://dreamprompting.com/api/v1, and POST to /api/v1/chat/completions. To target NIM specifically, pin a model id with the nvidia/ prefix. Here is a curl example that pins a small Llama NIM model:

curl https://dreamprompting.com/api/v1/chat/completions \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/meta/llama-3.1-8b-instruct",
    "messages": [
      {"role": "user", "content": "Explain NVIDIA NIM in one sentence."}
    ]
  }'

If you prefer Python, the official OpenAI SDK works without changes beyond the base URL and key. Set base_url to the gateway and api_key to your free key:

from openai import OpenAI

client = OpenAI(
    base_url="https://dreamprompting.com/api/v1",
    api_key="YOUR_API_KEY",
)

resp = client.chat.completions.create(
    model="nvidia/meta/llama-3.1-8b-instruct",
    messages=[
        {"role": "user", "content": "Give me three uses for a NIM model."}
    ],
)

print(resp.choices[0].message.content)

Want streaming tokens instead of one blocking response? Add "stream": true to the request body (or stream=True in the SDK) and read the chunks as they arrive. If you would rather not write code at all, you can try prompts right now in the browser with the hosted chat playground.

Why NVIDIA NIM feels slow, and what to do about it

The very large NIM models cold start. When a big model has been idle, the first request waits while capacity spins up, which can run to several seconds before a single token arrives. Smaller models like llama-3.1-8b-instruct stay responsive because they are kept warm by steady traffic. Three things help: pin a small model when latency matters more than capability, keep a periodic request going so your model of choice stays warm, and put failover in front of the call so a cold model does not become a timeout.

The gateway does the last of those for you. Its NVIDIA default is meta/llama-3.3-70b-instruct, chosen for answer quality rather than latency, and if a model is slow or unavailable it fails over automatically. Pin nvidia/meta/llama-3.1-8b-instruct yourself when you would rather have the fast small model. You can also send "model": "auto" and let it pick a healthy option. Limits are generous: 100 requests per minute per IP, plus a rolling 24 hour account quota of 5,000 requests and 500,000 tokens, which covers prototypes and light production traffic.

FAQ

Is NVIDIA NIM free? NVIDIA's own hosted endpoint gives new accounts a limited pool of free credits and is paid after that, and self-hosting a NIM container needs an NVIDIA AI Enterprise license plus your own GPU. Reaching the same models through this gateway is free with no card and no billing: sign in with Google, copy your key, send it as a Bearer token, and call the endpoint.

Is the endpoint OpenAI-compatible? Both are. NVIDIA's https://integrate.api.nvidia.com/v1 and this gateway's https://dreamprompting.com/api/v1 accept the standard chat completions body, so switching between them is a base_url change and nothing else. See the API reference for the full parameter list.

Can I run NIM models on my own GPU instead? Yes, if you have the VRAM. Most NIM models are open weights, so you can serve them yourself. Check what your card can hold with the VRAM calculator, or see which GPUs run a given model.

How do I know which NIM model ids are valid? Call GET /api/v1/models to see the live list. Model availability shifts over time, so read it fresh rather than assuming an id still exists.

Why did my request use a different model than I pinned? If a pinned NIM model is slow or unavailable, the gateway fails over to a healthy model so your call still succeeds. Use a specific id when you need determinism, or auto when you want speed and reliability.

Start building

Ready to reach NVIDIA NIM for free? Browse the live catalog on the models page to find the exact id you want, then wire it up with the full API documentation. One endpoint, one free key, automatic failover, and NIM models at your fingertips.