NVIDIA build.nvidia.com: 100+ AI Models via Free API (No More Local Models)

NVIDIA just removed the biggest barrier to local AI: the hardware. With build.nvidia.com, you get access to 100+ cutting-edge AI models (Llama, Nemotron, DeepSeek, GLM-4, gpt-oss…) via a free, OpenAI-compatible API. No more downloading 200 GB of weights, no more needing a $2,000 GPU, no more "disk full". Change two lines in your code and run on NVIDIA GPUs in the cloud. We break it all down, with real copy-pasteable code.

flowchart LR
    A["Your code / IDE"] --> B["OpenAI SDK"]
    B --> C["NVIDIA Endpoint
integrate.api.nvidia.com/v1"]
    C --> D["NIM · GPU cloud"]
    D --> E["100+ models
Llama · Nemotron · DeepSeek"]

What is build.nvidia.com (NVIDIA NIM)?

Since 2024, NVIDIA has been rolling out NIM (NVIDIA Inference Microservices): a layer that serves AI models optimized for its GPUs. There are two halves: containers you can self-host, and a hosted inference catalog — the latter is making waves today on build.nvidia.com.

The brilliant idea: the endpoint is compatible with the OpenAI API. In practice, any tool, framework or app that can point to a custom “base URL” can target NVIDIA with no other code changes. Streaming, function calling, tool use: everything works exactly like OpenAI.

Why this is a game-changer: no more local models

Running a large model at home is a slog: downloading tens or hundreds of gigabytes, having enough VRAM, managing CUDA drivers, and accepting that your laptop will heat up like a radiator. The cloud endpoint solves all of that in one go.

Get started in 3 minutes: your nvapi- key

The only prerequisite is a free NVIDIA Developer Program account. No credit card required.

Steps: (1) go to build.nvidia.com and sign in, (2) open any model, (3) click Get API Key, (4) copy the key that starts with nvapi- — it is shown only once. Then export it and install the SDK:

# 1. Store your key in an environment variable
export NVIDIA_API_KEY="nvapi-xxxxxxxxxxxxxxxxxxxxxxxx"

# 2. Install the OpenAI SDK (yes, the one from OpenAI)
pip install openai

Your first call (Python)

Here is the shortest working example. Notice the only two NVIDIA-specific things: base_url and the key.

from openai import OpenAI

client = OpenAI(
    base_url="https://integrate.api.nvidia.com/v1",
    api_key="nvapi-VOTRE_CLE",   # or os.environ["NVIDIA_API_KEY"]
)

resp = client.chat.completions.create(
    model="deepseek-ai/deepseek-v4-pro",
    messages=[{"role": "user", "content": "Explain RAG in 3 sentences."}],
    temperature=0.7,
    max_tokens=1024,
)
print(resp.choices[0].message.content)

RAG (Retrieval-Augmented Generation) combines document retrieval
with a language model. It first retrieves relevant passages from
a knowledge base, then injects them into the prompt so the model
answers with up-to-date facts instead of relying only on its memory.

Streaming, Node.js and cURL

Streaming (displaying the answer word by word) works exactly like OpenAI:

stream = client.chat.completions.create(
    model="meta/llama-3.3-70b-instruct",
    messages=[{"role": "user", "content": "Write a haiku about the GPU."}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

In Node.js, it is the exact same SDK:

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://integrate.api.nvidia.com/v1",
  apiKey: process.env.NVIDIA_API_KEY,
});

const r = await client.chat.completions.create({
  model: "openai/gpt-oss-120b",
  messages: [{ role: "user", content: "Hello from Node!" }],
});
console.log(r.choices[0].message.content);

And in plain cURL, with no SDK at all:

curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/nemotron-3-ultra-550b-a55b",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Connect Cursor, Claude Code and LangChain

Because the endpoint is OpenAI-compatible, you can plug it into any tool that accepts a custom base URL. For Cursor: in the model settings, enter the base URL https://integrate.api.nvidia.com/v1 and your nvapi- key — chat, autocomplete and the agent will switch to NVIDIA-hosted models. The same logic applies to Claude Code, LangChain, CrewAI or AutoGen: only change the base URL and model ID.

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="https://integrate.api.nvidia.com/v1",
    api_key="nvapi-VOTRE_CLE",
    model="meta/llama-3.3-70b-instruct",
)
print(llm.invoke("Give me 3 AI project ideas for beginners.").content)

curl -s https://integrate.api.nvidia.com/v1/models \
  -H "Authorization: Bearer $NVIDIA_API_KEY" | jq '.data[].id'

Available models + Agentic Skills

The catalog exceeds 100 models: Llama and Nemotron families (NVIDIA), DeepSeek, GLM-4 (Zhipu AI), gpt-oss (OpenAI open weights), MiniMax, plus embedding, speech recognition and simulation models. Each model page shows its pricing: some are completely free, others consume your credits.

NVIDIA also pushes Agentic Skills: ready-to-use capabilities your agent can call, organized by domain — AI & Machine Learning, Accelerated Computing, Physical AI, Developer Tools. Perfect for building an agent that acts, not just chats.

Free tier limits (to know)

The free tier is generous for prototyping, but has clear boundaries:

If you exceed the rate limit, the API returns a 429 Too Many Requests error. The classic workaround: space out calls and retry with exponential backoff.

import time
from openai import OpenAI

client = OpenAI(base_url="https://integrate.api.nvidia.com/v1", api_key="nvapi-VOTRE_CLE")

def chat_with_retry(messages, model="meta/llama-3.3-70b-instruct", attempts=5):
    for i in range(attempts):
        try:
            return client.chat.completions.create(model=model, messages=messages)
        except Exception as e:
            if "429" in str(e) and i < attempts - 1:
                time.sleep(2 ** i)   # 1s, 2s, 4s, 8s...
                continue
            raise

When to stay local anyway

The endpoint does not kill local inference — it complements it. Stay local when: (1) your data is sensitive and must not leave, (2) you work offline, (3) you need ultra-stable latency, or (4) you are at large scale and self-hosting becomes cheaper. For everything else — prototyping, demos, side projects, learning — the free endpoint is unbeatable. And when you want to move to production self-hosted, NVIDIA offers a free R&D license (up to 16 GPUs) and a 90-day AI Enterprise trial.

FAQ

📬 Want to receive this kind of guide every week? Subscribe for free — real code, zero fluff.