Nemotron 3 Ultra really is free and really is good. The free part runs out at 50 requests a day

NVIDIA released it on 4 June 2026. It uses a mixture of experts architecture: there are 550 billion parameters in total, but only 55 billion of them fire on any given request, not all of them. That is why it is fast for its size. The context window goes up to 1 million tokens. The license is OpenMDW 1.1 and it permits commercial use, which matters because most models labelled "open weight" do not.

What is extra below: the verified setup for both free routes, exactly where each one stops, the step in OpenCode's first run that walks you into handing over billing details, and why downloading the weights is not an option for very nearly anyone.

Two sentences circulate about this model and both are measurable, so I checked them. The first is "the biggest free model to date". It is not. By Hugging Face's own figure, Nemotron 3 Ultra carries 560 billion parameters. Kimi K3, sitting in the same place, carries 2.78 trillion, roughly five times as many, and it is downloadable too. There are larger ones over at DeepSeek and Thinking Machines as well. Size is not where Nemotron holds a record.

The second is "even better than Claude". NVIDIA makes no such claim. The sentence on its own research page and in the abstract of its technical report is this: up to 6 times higher inference throughput compared to open models, at comparable accuracy. The models it compares against are GLM, Kimi and Qwen, all open. They publish no comparison against Claude or GPT, and the absence of that comparison is itself informative.

The right expectation: among open models you are getting one of the best on speed and efficiency, and getting it for nothing. You are not getting something that replaces Claude. That is not bad news, because accuracy was never this model's selling point. Finishing the same job in fewer tokens and less time pays off directly when your daily allowance on the free tier is counted.

There are two separate free routes and people conflate them. The first is OpenRouter's free endpoint: the model name ends in ":free" and both input and output really do cost zero. The second is NVIDIA's own API: you open an account, it asks for no card, and you take your key. Different setup, different limits.

The second row is the most useful thing on this page. If you load 10 dollars of credit on OpenRouter once, your free tier's daily allowance goes from 50 to 1,000, and that 10 dollars stays on your account because free requests do not spend it. So a one time 10 dollars gets you twenty times more use out of this model.

The weights genuinely are public, you can start downloading them right now. Downloading is not the problem, running them is. NVIDIA's published version comes to roughly 1.1 TB, and the minimum hardware on the model card is 8 H200s or 16 H100s. Compressed versions bring that down, but not far enough.

The conclusion: if you are going to use this model, you are going to use it over the cloud. Both setups below do exactly that. The download route makes sense only if you have servers sitting around.

OpenCode is an open source coding agent that runs in your terminal. It is MIT licensed and you pick the model; the tool itself does not care which one you use. Installing it is a single command.

# macOS and Linux
curl -fsSL https://opencode.ai/install | bash

# Windows
scoop install opencode

# or anywhere, if you have Node
npm install -g opencode-ai

If you cannot see the free version of Nemotron 3 Ultra in the model list, or you would rather not pick it every time you start, you can pin it in the config file. That file is either opencode.json at your project root or .config/opencode/opencode.json in your user folder; the second one applies to every project.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "openrouter/nvidia/nemotron-3-ultra-550b-a55b:free",
  "provider": {
    "openrouter": {
      "models": {
        "nvidia/nemotron-3-ultra-550b-a55b:free": {}
      }
    }
  }
}

The ":free" on the end is not decoration, it is precisely what separates the paid endpoint from the free one. Delete it and the model still works, but now it spends your credit.

You can cut OpenRouter out entirely and connect straight to NVIDIA. Opening an account is free and asks for no card, and your key starts with nvapi. OpenCode does not know this provider out of the box, so you introduce it yourself using the OpenAI compatible provider pattern.

{
  "$schema": "https://opencode.ai/config.json",
  "model": "nim/nvidia/nemotron-3-ultra-550b-a55b",
  "provider": {
    "nim": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NVIDIA NIM",
      "options": {
        "baseURL": "https://integrate.api.nvidia.com/v1",
        "apiKey": "{env:NVIDIA_API_KEY}"
      },
      "models": {
        "nvidia/nemotron-3-ultra-550b-a55b": {
          "name": "Nemotron 3 Ultra"
        }
      }
    }
  }
}

The name "nim" here is yours to choose. If you name the provider "nvidia", the model address turns into something like nvidia/nvidia/nemotron-3-ultra-550b-a55b with nvidia written twice, which gets confusing, so a short name is easier. We read the key from an environment variable instead of writing it into the file; whatever you named that variable is what goes inside the braces.

The last one deserves a pause, because most of these videos start right there. The thing that became free is the model. What you are swapping is the agent, the layer that talks to you in the terminal and touches your files. There is another route that solves the same problem while keeping the Claude Code interface, and it sits at number 110 in this archive. Which one you pick comes down to this: do you want to keep the interface, or the model.