01What Nemotron 3.5 Super VL Is and How It Works

Nemotron 3.5 Super VL is a free AI vision-language model released by Nvidia on October 15, 2024. It reads screenshots, PDFs, charts, client reports, and video, then reasons about what it sees and takes action by calling tools or writing code. You can use it inside an AI agent stack so your agent no longer needs you to describe images in text first.

The model has 120 billion total parameters, but only 12 billion activate for each word it writes. Think of it as a team of 120 people where only the 12 best experts stand up for any single job. This design gives you the intelligence of a large model and the speed of a small one.

It outputs 630 tokens per second on a 100,000-token input, about 4.5 times faster than the original Nemotron 3 Super and 2.7 times faster than Nemotron 3 Ultra.

02Main Features of Nemotron 3.5 Super VL

  • Multi-modal vision: It reads text, images, documents, and video inside a single model.
  • Tool calling and code writing: After reading visual input, it can call external tools or write Python code to act on the information.
  • Distilled from Nemotron 3 Ultra: Nvidia used Ultra as a teacher model, had it solve thousands of problems, then trained Super VL to reproduce the same answers more efficiently.
  • Multi-token prediction: Instead of writing one word at a time, it guesses and checks several words ahead simultaneously, which speeds up output.
  • Lilikore tuning: The model is specifically tuned for DGX Spark and low concurrency workloads, so it doesn't waste compute when the job is light.
  • 1-million-token context window: You can feed it roughly 750,000 words at once, such as a full month of community questions or a long onboarding document.

03How to Set Up Nemotron 3.5 Super VL in Your Agent Stack

  1. Obtain the Agent OS zip file that contains Hermes Agent, OpenClaw, and OpenCode pre-configured together. You can get it inside the AI Profit Boardroom or from the AI Profit Boardroom website.
  2. Install the zip file on your system. Day one of the Agent OS roadmap covers exactly how to do this.
  3. Open Hermes Agent and locate the model drop-down menu.
  4. Select Nemotron 3.5 Super VL from the list of available models.
  5. Configure the model to be the default vision layer. Now, any task that involves an image, PDF, or screenshot will automatically be routed to Nemotron 3.5 Super VL for processing.
  6. Test the setup with a real visual task, such as asking it to read a screenshot of a workflow and suggest what to automate.

After these steps, your agent stack can read and act on visual data without manual intervention. The Agent OS keeps Hermes, OpenClaw, and OpenCode sharing context so information flows through the whole stack.

04How to Use Nemotron 3.5 Super VL with a Real Example

Here is a real example shown in the episode. Suppose a member of your community sends a screenshot of their current workflow and asks, "What should I automate next?" In the old setup, your AI agent would need you to describe that screenshot in text first. With Nemotron 3.5 Super VL loaded into Hermes Agent, the agent reads the image directly.

It figures out what is happening in the workflow and returns a real answer without any manual text work.

This removes the manual middle layer entirely. Inside the Agent OS, once Nemotron 3.5 Super VL reads something visual inside Hermes (a screenshot, a chart, a PDF), that information flows to OpenClaw so it can act on the data, and to OpenCode so it can build software based on what it saw. Everything stays connected.

The model also uses about 60% fewer output tokens per task compared to the predecessor Nemotron 3 Super, so it completes tasks faster.

05Benchmark Results and Comparisons

In the episode, Julian showed three key numbers. First, Nemotron 3.5 Super VL scored 63.44% on the Hermes benchmark, which beat Nvidia's own bigger model Nemotron 3 Ultra at 56.33%. The smaller, faster, free model came out ahead.

On LangChain behavioral evals, which test how an agent performs across a full workflow, Super VL scored 87.05% and Ultra scored 84.3%.

The model is 4.5 times faster than the original Nemotron 3 Super and 2.7 times faster than Nemotron 3 Ultra, producing 630 tokens per second on a 100,000-token input. It also uses 60% fewer output tokens per task compared to its predecessor. The reason it beats the larger model is distillation: Nvidia used Ultra as the teacher, had it solve thousands of problems, then trained Super VL to hit the same answers more efficiently, learning shortcuts the teacher had to discover the hard way.

06Tips, Limits, and Mistakes to Avoid

  • Use the right model for the right job: Nemotron 3.5 Super VL is designed for vision tasks. Do not use it for non-vision tasks where a text-only model would be more efficient. The Agent OS is built to route the correct model to each task automatically.
  • Lilikore tuning: This model is optimized for DGX Spark and low concurrency workloads. If you run it on hardware that does not match these specs, you may not get the full speed benefit.
  • Don't skip testing: After you slot the model into Hermes, run a real visual task to confirm it reads and acts correctly. The first test can use a screenshot of a workflow or a PDF.
  • Avoid token waste: Because the model uses multi-token prediction and reduces output tokens by 60%, always-on agents handling multiple tasks will compound those savings. But if your task does not involve visual input, you are wasting compute.
  • Check compatibility: Make sure your agent stack (Hermes, OpenClaw, OpenCode) is correctly configured to pass vision tasks to this model. The Agent OS zip file includes pre-configured settings.

07Who Nemotron 3.5 Super VL Is For and Why It Matters

This model is for business owners, AI developers, and agency operators who want to automate visual and document-heavy tasks. If your AI agents currently cannot read screenshots, PDFs, charts, or video, Nemotron 3.5 Super VL gives them vision without extra cost. It removes the manual step of describing images in text, so your agent can go straight from seeing to acting.

Right now most people are not running multi-model agent stacks that use vision models inside Hermes Agent. That creates a window of opportunity. The people who set up this system now will be operating at a different level in six months.

Inside the AI Profit Boardroom, members already run this exact stack with Nemotron 3.5 Super VL as their vision layer. It is live and working, not theoretical.