build.nvidia.com is NVIDIA's playground for AI models. Try hundreds of the leading open models — Llama, DeepSeek, Qwen, NVIDIA's own Nemotron, plus vision, speech, and biology models — right in the browser, get an OpenAI-compatible API to call them from your app, then deploy the exact same model on your own hardware when you're ready. Free credits to start. Here's the builder's map.
Not toy demos. You can prompt hundreds of leading open models — Llama, DeepSeek, Qwen, NVIDIA's Nemotron, plus vision, speech, and biology models — straight in the browser, no setup.
Every model comes with a ready endpoint. If your code already talks to OpenAI, you change the base URL and key and you're calling NVIDIA-hosted models — minimal rewrite.
Under the hood each model is a NIM — an NVIDIA Inference Microservice, a prepackaged, GPU-optimized container. The hosted API you're testing is just that NIM running on NVIDIA's own GPUs.
The part that matters: build against the free hosted API, then run the same NIM on your own cloud or GPUs in production. Same model, same behavior, no rewrite, no vendor lock to their infra.
Go to build.nvidia.com and browse by task — chat, reasoning, vision, RAG, speech, code. Filter down to the model you want to test.
Prompt the model right there. Browsing is open; a free NVIDIA account lets you run it. Check quality and latency before you write any code.
Each model page hands you copy-paste code in Python, Node, or curl, pre-filled with your key. It's an OpenAI-style call — the base_url just points at NVIDIA.
Because it's OpenAI-compatible, most SDKs work by swapping two values: the base URL and the key. Prototype for free on the included credits.
When you outgrow the hosted API, pull the NIM container and run it on your own GPUs or cloud — identical model, no rewrite. That's the whole point of the platform.
# OpenAI-compatible — you only change base_url + api_key
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1",
api_key="nvapi-...", # from build.nvidia.com
)
resp = client.chat.completions.create(
model="meta/llama-3.1-70b-instruct", # swap for any model in the catalog
messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)The playground and the API snippets are free to explore. Browse the models, try one in your browser, and grab the endpoint.