Nemotron 3.5 Super VL is NVIDIA's new open multimodal model for AI agents. It reads text, images, documents and video, then writes text back. It has 120B total parameters but only 12B active per token, it handles up to 1M tokens of context, and you can use it free right now through OpenRouter and OpenCode.
📺 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 tested it the day after NVIDIA released it on 15 October 2026. Here is everything you need to know: what it is, the numbers, what it can actually do, where to run it, and my honest take on whether it belongs in your stack.
What Nemotron 3.5 Super VL Actually Is
Think of it as one model that can see and act. Most open models are text only. You paste words in, you get words out. This one can look at a screenshot, a PDF, a chart or a short video clip and reason about it, then call tools and finish the job.
NVIDIA built it for agents. That means reasoning, instruction following and tool use all live inside the same model. You also get thinking and non-thinking modes, so you can switch reasoning on for hard problems and off when you just want a fast answer.
Under the hood it is a mixture-of-experts model with a hybrid architecture that blends Mamba layers with attention. It was distilled from Nemotron 3 Ultra, NVIDIA's bigger model. It also uses multi-token prediction, where an internal drafter proposes several tokens at once and the model checks them in a single pass. That is a big part of why it is quick.
If you want to learn how to plug models like this into a working agent setup, that is exactly what we do inside the AI Profit Boardroom. More on that later.
Nemotron 3.5 Super VL Key Numbers
Here are the specs in one place. Every figure comes from NVIDIA or the official model listings.
| Spec | Value |
|---|---|
| Total parameters | 120B |
| Active parameters per token | 12B |
| Architecture | Hybrid Mamba + attention, mixture-of-experts |
| Speed trick | Multi-token prediction (built-in drafter) |
| Inputs | Text, images, documents, video (sampled frames / short clips) |
| Output | Text |
| Context window | Up to 1M tokens (usable amount depends on memory and runtime) |
| Reasoning | Thinking and non-thinking modes |
| License | OpenMDW License Agreement v1.1 (open weights) |
| Release date | 15 October 2026 |
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
What It Can Do: Vision Plus Agents
The VL stands for vision-language. That is the whole point of this model. Here is where it earns its place:
- Screenshots. Show it a broken web page or an app screen and ask what is wrong.
- PDFs and documents. Feed it contracts, reports or invoices and pull out what matters.
- Charts. Ask it to read a graph and explain the trend. Chart reasoning is one of its strongest scores.
- Video. It works through sampled frames and short clips, so it can describe what happens in a recording.
- Tool use. It can call tools, write code and run multi-step tasks inside an agent harness.
So the simple rule: if the job starts with a picture or a file, this model is worth trying. If the job is pure text and needs frontier-level reasoning, you probably want something bigger.
Benchmarks Summary
All of these numbers are NVIDIA's own. Nobody has independently replicated them yet, so treat them as a starting point, not a final verdict.
- Hermes Bench: 63.44%, versus 56.33% for Nemotron 3 Ultra, per NVIDIA.
- LangChain evals: 87.05%, versus 84.3% for Ultra.
- CharXiv (chart reasoning): 80.8%, versus 63% for Nemotron 3 Nano Omni. That rises to 84% with Python code interpretation.
- OCRBench v2: among the top open models, NVIDIA reports.
- SWE-bench and Terminal-Bench: beats Nemotron 3 Super, though NVIDIA gave no numbers.
- Indic benchmarks: leads Sarvam-105B and Gemma-4-31B-It, again without numbers in the text.
The interesting bit is that a 12B-active model beats the bigger Ultra on agent benchmarks. That is the distillation and the agent training paying off. For the full breakdown, read the Nemotron 3.5 Super VL benchmark page. I also keep my own practical scores on Goldie Bench, where I test models on real business jobs.
📺 Watch: New FREE Nvidia Nemotron 3 Super + Ollama
Speed
Speed is where NVIDIA is pushing hardest. Here is what they report:
- Up to 5x higher single-GPU throughput on agentic tasks versus open models its size.
- Over 600 tokens per second on agentic tasks at 100K input tokens, on a Baseten endpoint.
- 4x faster task completion across 10,000 coding tasks on B200.
- On DGX Spark, LiLiCorr speculative decoding gives up to 2.3x faster generation.
- With Dynamo on 16 B200 GPUs at 50 tokens per second per user: up to 8x as many users and 85% lower time to first token.
EXO Labs ran an independent evaluation on a 128GB DGX Spark and put it on the accuracy versus task-time Pareto frontier. That is the closest thing to outside confirmation so far.
It also ships in NVFP4, a 4-bit format trained with quantization-aware distillation. On Blackwell GPUs it runs in true 4-bit. On older hardware the same checkpoint runs with 4-bit weights and 16-bit activations.
A Genuinely Open Release
NVIDIA did not just drop weights and walk away. The release includes:
- Base, post-trained and NVFP4-quantized checkpoints.
- Training recipes, data and RL environments.
- Open datasets on Hugging Face, including 40 vision-language datasets.
- A usage cookbook on GitHub.
That matters if you want to fine-tune it or build on it. You can see how it was made, not just use the end result.
Where to Use It: Hosted vs Local
This is not a normal-laptop model. Running it locally needs roughly 70 to 90GB of memory. Think a 128GB DGX Spark or a big-memory workstation.
Hosted (best for most people):
- OpenRouter lists it as nvidia/nemotron-3.5-super-vl, plus a :free variant.
- OpenCode lets you pick it free from the model drop-down.
- Clouds: Amazon SageMaker JumpStart, Microsoft Foundry, Baseten, Fireworks AI and Together AI.
If you just want to try it without spending a penny, the Nemotron 3.5 Super VL free guide walks through every free route.
Local (if you have the hardware):
- Ollama: ollama run nemotron-3.5-super (about 80-90GB).
- LM Studio: listed as nvidia/nemotron-3.5-super-vl, smallest GGUF is Q4_K_M at about 70GB.
- Weights on Hugging Face and ModelScope; also vLLM, SGLang, Unsloth and Exo.
Agent harnesses that support it include OpenCode, OpenClaw, Pi, LangChain, Cline, OpenHands and Hermes Agent.
My Test and Honest Take
I opened OpenCode inside my Agentic Operating System, picked Nemotron 3.5 Super VL from the model drop-down, and asked it to "build a beautiful to-do list app, make it entertaining, engaging." It built it, and the app worked first time.
I also use it inside Hermes Agent, where it shows up as one of the models you can pick. If you want that setup, the Nemotron 3.5 Super VL Hermes Agent walkthrough shows how.
Here is the honest bit. This is not a frontier model. It is not going to replace Opus 5.5. But it is free, it is open, it is fast, and it can read screenshots, PDFs, charts and invoices. So slot it in for jobs that start with a picture or a file, and keep your frontier model for the heavy thinking.
One tip: if you run out of free tokens in OpenCode, switch to another free model there, like Big Pickle, or move the job into Hermes.
Want help working out where a model like this fits in your business? Book a free AI strategy session and we will map it out together.
Who It Is For
- Business owners running agents who want a free model for document and screenshot jobs.
- Developers who want an open, fast coding and tool-use model they can self-host.
- Teams with a DGX Spark or big workstation who want private, local multimodal AI.
- Anyone on a budget who wants to stop paying for every single agent call.
It is not for you if you need the absolute best reasoning on hard text problems, or if you only have a normal laptop and refuse to use a hosted option.
My Agent OS runs Hermes, OpenClaw, OpenCode and Claude side by side, and you can swap new open models like this one straight in. The Agent OS guide explains how it all fits together. Inside the AI Profit Boardroom ($69/mo) you get the Agent OS zip, a video tutorial, a 30-day roadmap, updates, 4 weekly coaching calls and daily tutorials, alongside 3,400+ members.
FAQ
Is Nemotron 3.5 Super VL free?
Yes. The weights are open under the OpenMDW License v1.1, and you can use it free on OpenRouter's :free variant and inside OpenCode. Running it on your own hardware is also free, if you have around 70-90GB of memory.
Can I run it on my laptop?
Not a normal one. The smallest LM Studio build is about 70GB and the Ollama version is about 80-90GB. Most people should use it hosted instead.
Is it better than Opus 5.5?
No. It is not a frontier model. Its strength is being free, open, fast and good at reading images, documents and charts inside agent workflows.
What does the VL mean?
Vision-language. It takes text, images, documents and video frames as input and returns text, so it can see what you show it and then act on it.
If you want a plan for using free open models like this across your agents, grab a free AI strategy session. Nemotron 3.5 Super VL is one of the best free tools you can add to an agent stack right now.
