Nemotron Super VL is the short name people use for NVIDIA's Nemotron 3.5 Super VL: an open model that reads text, images, documents and video and writes text back. "Super" is the size tier and "VL" means vision-language. That's it. If you're here because you saw the short name in a model drop-down or a tweet and wondered what you were looking at, this page gives you the quick version.
📺 Watch: NVIDIA Nemotron 3.5 Super VL Is INSANE! (FREE AI Model)
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I've used it free inside OpenCode and inside Hermes Agent, and I'll tell you where it fits and where it doesn't. If you want the full setup I run, with Hermes, OpenClaw, OpenCode and Claude side by side, that's inside the AI Profit Boardroom.
What Nemotron Super VL Actually Means
The name is three parts. Each one tells you something.
Nemotron
Nemotron is NVIDIA's family of open models. The 3.5 generation is the latest one, released in October 2026. NVIDIA publishes the weights, the training recipes, the data and a cookbook on GitHub, so this isn't a locked-down API product. You can download it, run it, and build on it.
Super
Super is NVIDIA's mid-to-large tier. Think of the family as a ladder. Nano sits at the small end. Ultra sits at the top. Super is in between: big enough to be properly useful for agents, small enough that it doesn't need a data centre to run. NVIDIA says this model was distilled from Nemotron 3 Ultra, which means the bigger model helped teach it.
VL
VL stands for vision-language. A plain language model only reads text. A vision-language model can also look at pictures. You hand it a screenshot, a scanned PDF, a chart or a photo of a receipt, and it can read what's in there and reason about it. Per NVIDIA, Nemotron 3.5 Super VL handles text, images, documents and video (video through sampled frames or short clips). The output is always text.
One thing that trips people up: in Ollama, the library name is nemotron-3.5-super. Ollama describes it as NVIDIA's vision-language model for coding, reasoning, tool use and document understanding. So if you see "Nemotron 3.5 Super" in Ollama, it's the same vision model, just with the shorter label.
What a Vision-Language Model Does for Agents
Here's why this matters. Most real work doesn't start as clean text. It starts as a screenshot someone sent you, a PDF invoice, a chart in a slide deck, or a photo taken on a phone. A text-only agent can't touch any of that without a separate OCR step bolted on. A vision-language model reads it directly.
That means one model can look at the file, understand it, call a tool, and write the answer. NVIDIA built reasoning, instruction following and tool use into the same model, with thinking and non-thinking modes you can switch between.
6 Practical Examples
- Screenshot to fix. Paste a screenshot of a broken web page or an error message and ask the agent what's wrong and how to fix it. No retyping the error.
- PDF to summary. Drop in a contract, a report or a scanned document and ask for the key points, the dates and the numbers. It reads the layout, not just the text.
- Chart to answer. Give it a chart from a slide or a dashboard and ask what the trend is, which month was highest, or what changed. Chart reasoning is one of the areas NVIDIA highlights most.
- Receipt to spreadsheet row. Photograph a receipt and ask for the vendor, date, total and tax pulled out as a clean list you can paste into your books.
- OCR on messy images. Handwritten notes, whiteboard photos, signs, product labels. It reads the text so your agent can act on it.
- Design to code. Show it a mockup or a screenshot of an app you like and ask it to build something similar. Inside OpenCode, that turns a picture into a working starting point.
The pattern is simple: if the job starts with a picture or a file, this model is worth slotting in.
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
The 60-Second Version of the Specs
Here's everything you need, per NVIDIA's technical blog and the LM Studio and Ollama listings.
| Spec | What it is | Plain English |
|---|---|---|
| Total parameters | 120B | A large model by open-weight standards |
| Active per token | 12B | Mixture-of-experts: only a slice works on each word, so it's faster |
| Architecture | Hybrid Mamba + attention | Built to handle long inputs efficiently |
| Multi-token prediction | Internal drafter | Guesses several tokens ahead and checks them in one pass |
| Inputs | Text, images, documents, video | It can see, not just read |
| Output | Text | It won't make images |
| Context | Up to 1M tokens | Usable amount depends on your memory and runtime |
| Modes | Thinking and non-thinking | Turn reasoning up for hard jobs, down for quick ones |
| License | OpenMDW License Agreement v1.1 | Open weights |
| Local size | About 70-90GB | Not a normal-laptop model |
A few headline numbers NVIDIA reports: 63.44% on Hermes Bench against 56.33% for Nemotron 3 Ultra, and 80.8% on CharXiv chart reasoning (84% with Python code interpretation). On speed, NVIDIA says up to 5x higher single-GPU throughput on agentic tasks than open models its size. These are NVIDIA's own figures, and there's no independent replication of the full set yet. For the full breakdown, read the main Nemotron 3.5 Super VL guide.
📺 Watch: New FREE Nvidia Nemotron 3 Super + Ollama
Where You Can Use It
You've got three realistic options.
- Free in OpenCode. Open the model drop-down and pick Nemotron 3.5 Super VL. This is how I tested it. The step-by-step is in my Nemotron 3.5 Super VL OpenCode walkthrough.
- Hosted on OpenRouter. The model ID is nvidia/nemotron-3.5-super-vl, and there's a :free variant. Free tiers come with rate limits, so check OpenRouter for the current ones. It's also available on Baseten, Fireworks AI, Together AI, Amazon SageMaker JumpStart and Microsoft Foundry.
- Locally. In Ollama, run ollama run nemotron-3.5-super. In LM Studio, search nvidia/nemotron-3.5-super-vl and hit Get; the smallest GGUF listed is Q4_K_M at about 70GB. You need roughly 70GB of RAM plus extra for images and context, so think 128GB DGX Spark or a big-memory workstation.
Ollama also lets you launch it straight into an agent harness: ollama launch hermes --model nemotron-3.5-super, or swap hermes for opencode, claude or openclaw.
If you're not sure which route fits your business, book a free AI strategy session and we'll work it out.
What I Found When I Tried It
I picked it from the OpenCode drop-down inside my Agentic Operating System and asked it to build a beautiful to-do list app, and to make it entertaining and engaging. The build came out working. I've also got it running as one of the models I can pick in Hermes Agent.
Honest take: it's not a frontier model. It's not going to replace Opus 5.5 for the hardest work. But it's free, it's open, it's fast, and it can read screenshots, PDFs, charts and receipts. That combination is rare. So I use it for jobs that start with a picture or a file, and I keep a bigger model for the rest.
If you run out of free tokens in OpenCode, switch to another free model there (Big Pickle is one I use) or carry on in Hermes. The way I wire models in and out of my setup is in the Agent OS guide.
Super VL vs the Plain Super Model
People mix these up. Nemotron 3 Super was the earlier text-focused Super model. Nemotron 3.5 Super VL adds vision, and NVIDIA says it beats Nemotron 3 Super on SWE-bench and Terminal-Bench setups (no numbers given). And because Ollama labels the vision model nemotron-3.5-super, the "VL" sometimes drops off the name entirely. If you want the naming sorted once and for all, read Nemotron 3.5 Super explained.
NVIDIA also released smaller Nemotron 3.5 Bolt edge models alongside it, a 74B-A9B for 24GB systems and an E4B for 4GB edge hardware. If 70GB is out of reach, those are the ones to look at.
Don't trust any benchmark table on its own, including NVIDIA's. Run the model on your own tasks. That's why I built Goldie Bench: same real jobs, every model, side by side.
FAQ
What does Nemotron Super VL stand for?
It's NVIDIA's Nemotron 3.5 Super VL. Nemotron is the model family, Super is the mid-to-large size tier, and VL means vision-language, so it reads images and documents as well as text.
Is Nemotron Super VL free?
You can use it free in OpenCode and through OpenRouter's :free variant, with rate limits. The weights are open and free to download, but running them locally needs about 70-90GB of memory.
Can it run on my laptop?
Not on a normal one. The smallest local version is about 70GB. Most people should use it hosted through OpenCode or OpenRouter instead.
Should I use it instead of Claude?
Use it alongside Claude, not instead of it. It's strong for screenshots, PDFs, charts and receipts and it's free, but it's not a frontier model. Keep a top model for your hardest work.
Want the full Agent OS with Hermes, OpenClaw, OpenCode and Claude ready to go, plus a 30-day roadmap, 4 weekly coaching calls and daily tutorials? Join 3,400+ members inside the AI Profit Boardroom for $69/mo. Or if you'd rather talk it through first, grab a free AI strategy session. Either way, Nemotron Super VL is worth ten minutes of your time this week.
