NVIDIA's 550B open reasoning model, wired into OpenCode as your coding agent. No GPU. No subscription. Copy, paste, done.
You are not downloading a 550B model onto your laptop. That would need roughly 4× B200 GPUs. What actually happens: NVIDIA hosts the model and gives you free API access, and you install a small terminal app called OpenCode that talks to it.
So: OpenCode lives on your machine. Nemotron lives on NVIDIA's servers. Your API key connects the two.
Open Terminal (Cmd + Space → type "Terminal") and paste this:
curl -fsSL https://opencode.ai/install | bash
Takes about 30 seconds. You'll see a progress bar and a green OpenCode logo when it's done.
Windows has no one-line curl installer. Pick one of these three. Easiest first:
Option A — winget (built into Windows 11 and updated Windows 10). Open PowerShell from the Start menu:
winget install SST.opencode
Option B — npm. Install Node.js first (LTS version, next-next-finish), then reopen PowerShell:
npm install -g opencode-ai@latest
Option C — Scoop, if you already use it:
scoop bucket add extras scoop install extras/opencode
OpenCode's interface needs proper colour support. Windows Terminal ships with Windows 11 and is a free install from the Microsoft Store on Windows 10. PowerShell inside Windows Terminal is the setup that behaves.
The installer usually prints "No config file found for zsh". That means your Mac doesn't know where OpenCode lives yet. Paste both lines:
echo 'export PATH=$HOME/.opencode/bin:$PATH' >> ~/.zshrc source ~/.zshrc
Then confirm:
opencode --version
A version number means you're good. command not found means the PATH line didn't save — quit Terminal, reopen, try again.
Close PowerShell and open a new window first. Windows only picks up new commands in a fresh terminal — this single step fixes most "it didn't work" messages. Then:
opencode --version
A version number means you're set.
Reopen PowerShell once more — installers add to PATH only for new sessions. Still nothing? Restart the PC, or reinstall using Option B (npm), which registers the command more reliably.
PowerShell's default script policy. Run this once, then reopen PowerShell:
Set-ExecutionPolicy -Scope CurrentUser RemoteSigned
Same on both platforms — and this is where most people get stuck, so read this bit properly.
If the page says you can't create keys, you're signed into a company or SSO organisation where an admin controls keys. Click the green Switch Org button and pick your personal org (usually your own name or "Default"). If nothing usable shows up, sign out and create a fresh account with a personal email.
The code goes into the same browser tab where you started signing in — nothing gets typed into a terminal.
The "Get API Key" button on a model page gives a testing key that expires in a few hours. Fine for a demo, useless the next morning. Generate a proper one from the API Keys page.
Move into any project folder and launch OpenCode:
cd ~/Desktop/my-project opencode
cd $HOME\Desktop\my-project opencode
Quick trick: open the folder in File Explorer, right-click empty space, choose Open in Terminal. No typing paths.
Inside the OpenCode screen, type:
/connect
Choose NVIDIA from the list, paste your nvapi- key, press Enter. The key is stored locally — you only do this once.
Still inside OpenCode:
/models
Type nemotron to filter, then pick Nemotron 3 Ultra 550B A55B and hit Enter. Or set it directly:
/model nvidia/nemotron-3-ultra-550b-a55b
Now just talk to it. Try: "read this folder and explain what the project does". It reads files, writes code, runs commands and fixes its own errors.
If you just want the model inside your own script, the API is OpenAI-compatible. Save your key as an environment variable first:
echo 'export NVIDIA_API_KEY="nvapi-your-key-here"' >> ~/.zshrc source ~/.zshrc pip install openai
setx NVIDIA_API_KEY "nvapi-your-key-here" pip install openai
setx saves it permanently, but only new terminals see it — close PowerShell and reopen before running your script.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://integrate.api.nvidia.com/v1",
api_key=os.environ["NVIDIA_API_KEY"],
)
stream = client.chat.completions.create(
model="nvidia/nemotron-3-ultra-550b-a55b",
messages=[{"role": "user", "content": "Write a Python script that renames every file in a folder to lowercase."}],
temperature=1,
top_p=0.95,
max_tokens=16384,
extra_body={"chat_template_kwargs": {"enable_thinking": True}},
stream=True,
)
for chunk in stream:
if chunk.choices and chunk.choices[0].delta.content:
print(chunk.choices[0].delta.content, end="")
| What you see | What to do |
|---|---|
| command not found: opencode | PATH line didn't save. Redo Step 02, then quit and reopen Terminal. |
| 'opencode' is not recognized | Open a brand-new PowerShell window. Still failing? Restart, or reinstall via npm. |
| running scripts is disabled | Run Set-ExecutionPolicy -Scope CurrentUser RemoteSigned, then reopen PowerShell. |
| npm not recognized | Node.js isn't installed, or you didn't reopen the terminal after installing it. |
| EPERM / permission errors | Turn on Developer Mode in Windows Settings, or run PowerShell as administrator once for the install. |
| 401 / Unauthorized | Key is wrong, expired, or has a stray space. Generate a fresh one from the API Keys page. |
| No API Key Permissions | Wrong organisation. Hit Switch Org, or sign up again with a personal email. |
| Nemotron not in /models list | Provider isn't connected. Run /connect again and pick NVIDIA. |
| Rate limited / 429 | Free tier throttle. Wait a minute, or switch to Nemotron 3 Super for lighter tasks. |
| Replies feel slow | It's a reasoning model — it thinks before answering. Long tasks are where it earns its keep. |
Built for long-running agent work — planning, coding, tool calling, debugging across big codebases. It's the strongest open model out of a US lab, though not the highest-scoring open model overall. Worth knowing before you argue about it in the comments.