There's a website that quietly hands developers free API access to 100+ AI models - Kimi, MiniMax, DeepSeek, GLM, Qwen, Mistral, Google's Gemma, OpenAI's open GPT-OSS models, and NVIDIA's own Nemotron series - all through one OpenAI-compatible endpoint, with no credit card required.
That website is NVIDIA Build (build.nvidia.com), and it runs on NVIDIA's NIM platform (NVIDIA Inference Microservices). The models are hosted on NVIDIA's own DGX Cloud, so you don't need a GPU, a subscription, or a paid plan to start building.
This is the complete setup guide: create your free key, test models in the Playground, drop the API into your own app, wire it into Cursor, connect it to Claude Code the correct way, and install NVIDIA's official agent Skills. Let's go.
What is NVIDIA NIM (and the NVIDIA free API)?
NIM stands for NVIDIA Inference Microservices - pre-packaged, GPU-optimized microservices that run large language models on NVIDIA hardware. Until mid-2024 this was an enterprise-only product you installed on your own servers.
Then NVIDIA opened a hosted model catalog on build.nvidia.com. Anyone in the free NVIDIA Developer Program can generate an API key and call those hosted models over the internet. The key thing that makes it so easy to use: every model is served through an OpenAI-compatible API. If your code, tool, or app already talks to OpenAI, switching to NVIDIA is usually a one-line change - you just swap the base URL and the key.
In plain terms: you get a real API key, pick a model, send requests, and pay nothing. No trial timer, no card on file.
What you get for free
The catalog holds 100+ models spanning text, code, reasoning, vision, embeddings, speech, and more. Names you'll actually recognize include:
| Model (examples) | Good for |
|---|---|
| Kimi K2 (Moonshot AI) | Long context, agentic tasks |
| DeepSeek V3.2 | Code + chat |
| GLM 4.7 / GLM (Z.ai) | Multilingual, coding |
| MiniMax M2.7 | Reasoning |
| Qwen3 Coder | Coding |
| GPT-OSS 120B (OpenAI open weights) | General purpose |
| Nemotron (NVIDIA) | Reasoning + agents |
Model names change often. New models are added almost weekly and IDs get versioned. Always copy the exact model ID from the model's page (see Step 3) instead of trusting a hardcoded name from any tutorial - including this one. The table above is illustrative, not gospel.
The free-tier limits (read this before you build)
At the time of writing, the free tier gives you roughly:
- ~1,000 free inference credits on signup (more available on request through NVIDIA's developer forum).
- A rate limit of about 40 requests per minute.
- No credit card and no per-token billing.
Some smaller models stay usable after your credits run out thanks to base daily quotas, while flagship models will return an error once credits are exhausted. These numbers change - NVIDIA tunes them over time - so check your live usage on the dashboard rather than assuming.
Important honesty check: the free hosted tier is meant for prototyping, learning, testing, and evaluation - not high-traffic production. Serving real end users at scale falls under NVIDIA's production terms, which means self-hosting NIM containers or an NVIDIA AI Enterprise license. Build and experiment freely; just don't design a live product around the free endpoint.
1 Create a free NVIDIA account
- Go to build.nvidia.com.
- Click Sign In / Sign Up and register with any email (personal or work - both are fine).
- Confirm your email. Joining the NVIDIA Developer Program is free and does not ask for a credit card.
Heads-up: some accounts are asked to verify a phone number via SMS before they can generate a key. If you don't get the code, try a different number.
2 Generate your free API key
- Open build.nvidia.com/settings/api-keys - or open any model in the catalog and click Get API Key.
- Click Generate Key.
- Click Copy Key and save it immediately - you only see it once.
Your key starts with nvapi- and is roughly 56 characters long. Treat it like a password: never commit it to Git or paste it publicly. Store it as an environment variable:
export NVIDIA_API_KEY="nvapi-XXXXXXXXXXXXXXXXXXXXXXXX"Don't truncate the key. If you copy a shortened version (with a …), you'll get 401 Unauthorized errors later. It must be the complete string.
3 Test any model in the built-in Playground
Every model page on build.nvidia.com has an interactive Playground - a chat box where you can test prompts before writing a single line of code.
Try it: open the GLM (or any) model page, type a prompt, and watch it respond. For example, ask it to "Create an Age Calculator using HTML, CSS, and JavaScript" - and it generates ready-to-use code in seconds.
Two things the model page gives you that you'll need next:
- A "View Code" button that shows a ready-made Python / Node / cURL snippet.
- The exact model ID inside that snippet (e.g.
z-ai/glm4.7). Copy this - it's what you'll paste into your own project.
4 Use the API in your own project
This is where the OpenAI-compatible design shines. The base URL for every model is:
If you're already using the OpenAI SDK, you change two lines - the base URL and the key.
Python
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1",
api_key="nvapi-XXXXXXXX", # or read from os.environ
)
response = client.chat.completions.create(
model="deepseek-ai/deepseek-v3.2", # copy the exact ID from the model page
messages=[
{"role": "user", "content": "Create an age calculator in HTML, CSS and JS."}
],
)
print(response.choices[0].message.content)JavaScript / Node
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://integrate.api.nvidia.com/v1",
apiKey: process.env.NVIDIA_API_KEY,
});
const response = await client.chat.completions.create({
model: "z-ai/glm4.7", // copy the exact ID from the model page
messages: [{ role: "user", content: "Hello!" }],
});
console.log(response.choices[0].message.content);cURL
curl https://integrate.api.nvidia.com/v1/chat/completions \
-H "Authorization: Bearer $NVIDIA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "z-ai/glm4.7",
"messages": [{"role": "user", "content": "Hello!"}]
}'The same key works across every model - just change the model value.
5 Connect NVIDIA models to Cursor
Cursor supports OpenAI-compatible endpoints directly, so this is a clean, two-minute setup:
- In Cursor, open Settings → Models.
- Enable the OpenAI API Key toggle and turn on Override Base URL.
- Paste the base URL:
https://integrate.api.nvidia.com/v1 - Paste your
nvapi-key. - Add a custom model using the exact NVIDIA model ID (e.g.
moonshotai/kimi-k2.6). - Save, then select your custom model in the chat/composer.
That's it - Cursor now runs on free NVIDIA-hosted models.
6 Connect NVIDIA models to Claude Code (the correct way)
Important correction to a common myth: you cannot paste the NVIDIA base URL straight into Claude Code the way you do with Cursor. Cursor speaks the OpenAI chat-completions format, but Claude Code speaks Anthropic's Messages API format. The free hosted NVIDIA endpoint is OpenAI-style, so Claude Code needs a small translation proxy in between. Skip this and you'll hit errors.
You have two easy paths:
Option A - Use a ready-made proxy (easiest)
Open-source tools like free-claude-code or claude-nvidia-proxy spin up a local server that translates OpenAI ⇄ Anthropic on the fly. You paste your nvapi- key into the proxy's config, start it, and point Claude Code at it:
export ANTHROPIC_BASE_URL="http://localhost:8082" # your proxy's address
export ANTHROPIC_AUTH_TOKEN="dummy" # proxy ignores this; it uses your nvapi- key
claudeThen map Claude Code's Opus/Sonnet/Haiku slots to NVIDIA model IDs (e.g. Kimi, GLM, DeepSeek) inside the proxy config.
Option B - Run LiteLLM as the proxy
Point Claude Code at a LiteLLM instance using nvidia_nim/-prefixed model names and set drop_params: true (Claude Code sends a few Anthropic-only parameters that NIM models don't understand). LiteLLM auto-routes to integrate.api.nvidia.com/v1.
Reality check: you'll be talking to open-source models, not Claude. Basic chat, code generation, and single-file edits work well. Extended thinking, long tool-use chains, and heavy multi-file edits can be hit-or-miss depending on the model you pick.
7 Install NVIDIA's official agent Skills
On build.nvidia.com/skills (the Skills tab), NVIDIA publishes verified "agent skills" - portable instruction sets (the same SKILL.md format used by Claude Code, Codex, and Cursor) that teach your AI agent how to correctly use NVIDIA's own stack: CUDA-X libraries, cuOpt, NeMo, RAG / AI-Q, DeepStream, Omniverse, and more. Every skill is signed and security-reviewed.
What they are (and aren't): these are for working with NVIDIA's platform and tools, not generic "make my agent smarter at everything" add-ons. If you're building on NVIDIA's ecosystem, they're gold; if you just want a free chat model, you can skip this step.
Prerequisites: Node.js and npm installed (that's what provides npx), plus an Agent-Skills-compatible client (Claude Code, Codex, Cursor, Kiro, and others).
Run these in your project's terminal:
# Browse the available NVIDIA skills first
npx skills add nvidia/skills --list
# Install interactively (pick from a menu)
npx skills add nvidia/skills
# Or install one specific skill into specific agents
npx skills add nvidia/skills \
--skill cuopt-numerical-optimization-api-python \
--agent claude-code \
--agent codex \
--agent cursorSmall clarification: you run the npx command in your terminal (inside your project), not by pasting it as a chat message. Claude Code and Codex can run terminal commands for you, but the command itself belongs in the shell.
Is it really free? Limits, rules & gotchas
- ✅No credit card, no per-token bill. Genuinely free to start.
- ⏳~40 requests/minute rate limit - fine for prototyping, not for a busy app.
- 🎟️Credit budget - ~1,000 free credits to start; request more via the forum. Flagship models stop once credits run out.
- 🧪Prototyping license - meant for dev, test, research, and evaluation. Production = self-hosted NIM or NVIDIA AI Enterprise.
- 🔁Model IDs drift - always copy the current ID from the model page.
- 🔒Guard your key - it's shown once; store it as an env var; never commit it.
Troubleshooting common errors
| Error | Likely cause | Fix |
|---|---|---|
401 Unauthorized | Key truncated or wrong | Re-copy the full nvapi- key |
402 Payment/limit | Free credits exhausted | Switch to a smaller model or request more credits |
404 Not Found (in Claude Code) | Anthropic model aliases still pointing at Anthropic IDs | Map Opus/Sonnet/Haiku to real NIM model IDs in your proxy |
429 Too Many Requests | Over 40 RPM | Slow down / add retry with backoff |
| JSON parse errors in Claude Code | Proxy not translating cleanly, or model returns bad tool calls | Try another model or set drop_params/thinking flags correctly |
FAQ
Is the NVIDIA API really free?
Yes - for development and testing. You join the free NVIDIA Developer Program (email only, no card) and generate an nvapi- key that unlocks the hosted model catalog. Production use has separate licensing.
Do I need a credit card to get an NVIDIA API key?
No. A valid email (and sometimes SMS phone verification) is all that's required.
What's the NVIDIA API base URL?
https://integrate.api.nvidia.com/v1 - it's OpenAI-compatible, so most OpenAI client code works with a two-line change.
Can I use NVIDIA's free API in Claude Code?
Yes, but not directly. Claude Code uses Anthropic's Messages format, so you need a translation proxy (e.g. free-claude-code or LiteLLM) between Claude Code and the OpenAI-style NVIDIA endpoint.
Can I use it in Cursor?
Yes, directly. Enable the OpenAI key override in Settings → Models, paste the base URL and key, and add the model ID.
How many requests can I make?
Around 40 requests per minute on the free tier, with a starting credit budget. Limits can change, so check your dashboard.
Which models are available?
100+, including Kimi, DeepSeek, GLM, MiniMax, Qwen, Mistral, Gemma, GPT-OSS, and NVIDIA Nemotron - plus vision, speech, and embedding models.
Final thoughts
For indie developers, students, and anyone building with AI on a budget, NVIDIA's free API is one of the most underrated tools available right now. You get frontier open-weight models, an OpenAI-compatible endpoint that drops into almost anything, a Playground to test ideas instantly, and official agent Skills for NVIDIA's stack - all without a credit card.
Grab your key, start in the Playground, then wire it into your favorite tool. Just remember the two rules that keep this guide mistake-free: copy the exact model ID from the model page, and use a proxy for Claude Code.