NVIDIA Gives You 45 AI Models Free. We Raced 4 of Them to Build a 3D Website

A friend is getting married and wants a little invitation website, but you can't code and every website builder wants a monthly plan. NVIDIA gives away 45 AI models for free, so we gave the same job to four of the best ones. Three of them failed. NVIDIA's own model, Nemotron 3.5 Super, built a working 3D site with gold rings you can spin, in under three minutes, for zero dollars.

Here's how to get the models, which ones actually answer, and exactly what happened in the race.

The job: a wedding invite site

Say a friend asks for a one-page site with their names, the date, a countdown, the schedule and a way to RSVP. That's exactly the kind of thing these free models are supposed to do. We added one stretch goal: two gold rings in real 3D that a visitor can grab and spin, right in the browser.

45 free models, 17 chatbots

The models live at build.nvidia.com. Open the models page, and on the left there's a filter called Free Endpoint. Tick that box, and when we tested it showed 45 models.

Not all of these are chatbots. Some clean up background noise, some track a body in 3D, some help self-driving cars. Seventeen are the kind you can talk to and ask for code.

The OK test: which ones answer

Before building anything, we sent all seventeen the same tiny message: reply with the word OK. A model that can't do that won't build you a website.

Thirteen answered. Ten of them came back in under ten seconds, and the fastest came back in two tenths of a second. Four didn't answer: one came back not found, and three just hung for 90 seconds, including two famous names people often recommend.

Keep this list. It saves you from waiting on a model that's stuck.

Get your free NVIDIA key

An API key is a private password that lets a tool use the models on your behalf. To get one:

1. Sign in, or make a free NVIDIA account. A lot of people now get asked to verify a phone number, and you may be asked to set up an NVIDIA Cloud account. Both are free.
2. Click API Keys, then Generate.
3. Give it a name and pick an expiry. Twelve months is a good choice.

NVIDIA shows the key only once, so copy it somewhere safe, and never paste it into a comment or a screenshot.

One more thing. NVIDIA says what you send to these free models is recorded to improve its products. So use them for things like this invite, never for passwords, client files or anything private.

Install OpenCode and connect NVIDIA

Next, the tool that does the building. OpenCode is a free coding agent that runs in your terminal, the text window where you type commands. You describe what you want, and it writes and saves the files.

• Mac or Linux: it's one line, from the OpenCode docs.
• Windows: you can use npm. A lot of videos show the wrong package name. The right command is npm install -g opencode-ai. If PowerShell shows a red wall saying scripts are disabled, there's a one-line fix you run once, then try again.

Then type opencode, then /connect, search for NVIDIA and paste your key. The free models show up in the list. A couple of the newest ones, MiniMax M3.1 Flash and Mistral Large 4, weren't in the list yet, so we added them with three lines in a small settings file. That file is in the free guide.

The race: 4 free models, 1 website

Same folder, same request, word for word, for every model: a one-page 3D wedding invitation for Maya and Sam, June 14, with gold rings you can drag to spin, a live countdown, a schedule and an RSVP form.

Laguna XS 2.1: fast, not finished

The fastest model on the list was done in 39 seconds. The page opens and looks clean. But there are no rings, and the countdown just shows dashes. One stray bracket in the code broke everything. It also put the wedding in 2024.

Mistral Large 4: Bad Request

The giant, with a trillion parameters in total (parameters are roughly the size of a model's brain). It worked for five and a half minutes, and then OpenCode showed one line: Bad Request. No file.

MiniMax M3.1 Flash: Too Many Requests

It got Too Many Requests on its very first message. We tried again later, by itself, and got the same thing. When a free model is that busy, it just isn't there for you that day.

Nemotron 3.5 Super: the winner

The model made by NVIDIA itself took 84 seconds. Two gold rings, lit and shining, and they really spin when you drag them.

One follow-up prompt fixes it

It wasn't perfect, though. The names were dark on a dark background, the rings sat on top of the text, and the countdown said zero, because June 14 had already passed this year.

So we did what you'd do with a person: told it all three things in one message. Seventy-eight seconds later, the names were gold and easy to read, the rings had moved to the side, and the countdown was live: 224 days to go.

We gave Laguna the same chance. It fixed the bracket in twelve seconds, but the 3D still didn't load. Fast isn't the same as finished.

The scoreboard: one model built it, in two prompts and under three minutes. Total spent, zero dollars.

The catch: honest rules

• Rate limits are real. When we ran three builds at once, we hit Too Many Requests right away. Run one at a time.
• It's for testing. NVIDIA says this free access is for building and testing, not for running a business on.
• Free models come and go. If one doesn't answer, switch to one from the fast list above.
• It's not private. Inputs and outputs are recorded, so keep personal data out.

Get the Rings Kit

Maya and Sam's invite is done, and it's theirs to keep, without writing a single line of code. If you want the exact prompt we used, the three-line OpenCode file and the list of models that actually answered, comment RINGS on the video, or grab it below.

How we tested: November 1 to 2, 2026, on one free NVIDIA key. Each chat model got one "Reply with the word OK" message with a 90-second limit. The website race ran in OpenCode, same folder and same prompt for every model, one fix prompt allowed. Free models change often. Independent, not affiliated with NVIDIA or OpenCode.