01What Is Nemotron 3.5 Super VL and How It Works

Nemotron 3.5 Super VL is a free multimodal AI model from Nvidia. It reads images, screenshots, PDFs, charts, and performs tasks from that visual input. This means you can give it a picture of a graph or a document and ask it to extract data, search the web, or write code.

The model has 120 billion total parameters, but for each word it writes, only 12 billion of them turn on. This trick makes it fast while keeping the brain of a much larger model. Unlike older models that need a separate text model and a separate vision model, this one combines both into a single system.

02Main Features and What’s New

The biggest new feature is that Nemotron 3.5 Super VL understands both text and images in one go. It handles screenshots, PDFs, charts, and photos. This makes it agentic — it can read a client report and then act on it without extra steps.

  • High speed: On one test with a 100,000-token input, it wrote over 600 tokens per second. That is roughly 100 words per second while holding a huge context.
  • Small footprint: Only 12 billion parameters activate per token. This makes it runnable on a desktop machine like the Nvidia DGX Spark.
  • Open source: Nvidia released the model weights, training data, training recipes, and 40 vision datasets. You can fine-tune it for your own use case.
  • Multi-token prediction: The model guesses several words ahead at once, then checks them together — like typing with autocomplete instead of one letter at a time.

03How to Set Up Nemotron 3.5 Super VL for Free

You can start using Nemotron 3.5 Super VL today without paying anything. Follow these steps to get it running inside a free tool called Open Code.

  1. Go to the Open Code website (open code is a free coding agent interface).
  2. Open the model dropdown menu near the input area.
  3. Select Nemotron 3.5 Super VL from the list of available models.
  4. Type your task — for example: 'Build a beautiful to-do list that is entertaining and engaging.'
  5. Press enter or click the run button. The model will start generating code or text based on your request.

You can also use it inside Hermes Agent, an agentic operating system. In Hermes Agent, you just pick Nemotron 3.5 Super VL from the model options and start giving it tasks that involve images or files.

04How to Use It: Walking Through Examples

In this episode, Julian Goldie shows Nemotron 3.5 Super VL inside Open Code. He says: 'Just build a beautiful to-do list, make it entertaining, engaging, et cetera.' The model then builds the list directly in the interface. You can see the result appear quickly.

Inside Hermes Agent, the model works with three other models. You can switch between them depending on the job. For tasks that start with a picture or file — like reading a screenshot of a client report or analyzing a receipt photo — you pick Nemotron 3.5 Super VL.

It reads the image and then acts on it, such as extracting data or searching the web.

You can also chain it with other models. If you run out of tokens on Nemotron 3 Super, you can switch to CatJam or Big Pickle inside Open Code. The model works with Open Router, so any agent that lets you pick a model can use it.

05Results, Tests, and Comparisons

Nvidia tested Nemotron 3.5 Super VL against other models and got strong numbers. On Hermes Bench, which tests agent tasks, the model scored 63.44%. Nvidia's own bigger model, Nemotron 3 Ultra, scored only 56.33%.

  • Lang Chain agent test: Super VL scored 87.05%, while Ultra scored 84.3%. The smaller model beat the larger one.
  • Chart understanding (Char XIV): Super VL scored 80.8%. Nvidia's previous small model, Nano Omni, scored 63%. When Super VL runs code to help, it hits 84%.
  • OCR Bench V2: The model near the top of open models for reading text inside pictures, like receipts.
  • Speed: On 10,000 tasks, Super VL finished up to four times faster than other open models in its class.

06Tips, Limits, and Mistakes to Avoid

Nemotron 3.5 Super VL is not a frontier model like Opus 5 or GPT-5. Do not expect it to beat top closed models in every test. It is a fast, free, open model built for agent work that involves images.

To use it locally, you need a powerful machine like an Nvidia DGX Spark or larger setup. It will not run well on weak laptops. If your task is purely text with no images, a smaller text model may be more efficient.

Avoid giving it vague instructions without examples. The model works best when you specify the visual input clearly — for example, 'Read this chart and tell me the trend' rather than just 'Analyze this.'

If you use Open Code, keep in mind token limits. You can switch to another model when you run out of tokens on Nemotron 3.5 Super VL.

07Who It Is For and Why It Matters

Nemotron 3.5 Super VL is for anyone who needs a fast, free AI that can see and act. Developers building agent systems will find it useful because it handles both text and vision in one model. Business owners who need to process receipts, client reports, or screenshots can use it to automate tasks.

It matters because it brings high performance to small setups. You no longer need expensive cloud models to read images and execute actions. With a DGX Spark desktop computer, you can run it locally and keep your data private.

Open source release also means you can customize it for your niche.

Julian Goldie uses it inside his agentic operating system alongside Claude, Hermes, and Openclaw. The model slots in for jobs that start with a picture or file, saving time and cutting costs.