NVIDIA Nemotron 3.5 Super VL is the open multimodal agent model NVIDIA shipped on 15 October 2026, and the business story is bigger than the model itself. NVIDIA released the weights, the training data and the recipes, then lined up NIM, Dynamo and the major clouds so companies can deploy it on day one.
📺 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 want to look at this from the business side. What did NVIDIA actually ship, why give so much away, how can you deploy it, and what do the customer results mean if you run AI agents in your company?
What NVIDIA Shipped on 15 October 2026
The headline is a 120B-parameter mixture-of-experts model with 12B active per token. It understands text, images, documents and video, and it writes text. Reasoning, instruction following and tool use all sit inside one model, with thinking and non-thinking modes you can switch between.
But the release package is the real news. NVIDIA put out:
- Checkpoints: base, post-trained and NVFP4-quantized versions.
- Training recipes, data and RL environments.
- Open datasets on Hugging Face, including 40 vision-language datasets.
- A usage cookbook on GitHub.
- A license: the OpenMDW License Agreement v1.1, with open weights.
It was distilled from Nemotron 3 Ultra and uses a hybrid Mamba plus attention design with multi-token prediction for speed. Context goes up to 1M tokens, depending on memory and runtime. For a plain-English rundown of the model itself, see the main Nemotron 3.5 Super VL guide.
If you want to actually use models like this in your business rather than just read about them, that is what we cover inside the AI Profit Boardroom.
Why NVIDIA Open-Sourced Weights, Data and Recipes
Most AI labs guard their training data and methods closely. NVIDIA handed them out. Here is why that makes sense for NVIDIA, reading the release for what it is.
NVIDIA sells the hardware. Every company that runs this model needs GPUs. The more businesses build agents on open models, the more compute they buy. Giving away the model grows the market for what NVIDIA really sells.
It is tuned for NVIDIA's stack. The NVFP4 format runs in true 4-bit on Blackwell GPUs, thanks to quantization-aware distillation. The same checkpoint still runs on non-Blackwell hardware with 4-bit weights and 16-bit activations, but the best numbers land on NVIDIA's newest chips.
Companies can customise it. With the recipes, data and RL environments, a business can fine-tune the model for its own workflows instead of starting from scratch. That makes Nemotron a foundation people build on, not just a model they try once.
It builds trust. Open data and recipes let teams see how the model was made. For companies that care about what goes into their AI, that counts.
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
NVIDIA Nemotron 3.5 Super VL Deployment Options
NVIDIA made sure there is a route for almost every kind of company.
| Route | What it is | Good fit for |
|---|---|---|
| NVIDIA NIM | NVIDIA's packaged inference recipes | Teams on NVIDIA infrastructure wanting a supported path |
| NVIDIA Dynamo | Serving recipes for scaling to many users | High-traffic agent products |
| Clouds | Amazon SageMaker JumpStart, Microsoft Foundry, Baseten, Fireworks AI, Together AI | Businesses already on a cloud provider |
| OpenRouter | nvidia/nemotron-3.5-super-vl, plus a :free variant | Quick testing and low-volume use |
| Self-hosted | Weights on Hugging Face and ModelScope; Ollama, LM Studio, vLLM, SGLang, Unsloth, Exo | Teams that need data to stay in-house |
The weights live on Hugging Face as nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. The Nemotron 3.5 Super VL Hugging Face page covers that route in detail.
For self-hosting, plan for around 70-90GB of memory. A 128GB DGX Spark is one of the cleanest ways to do it on a desk. EXO Labs ran an independent evaluation on a 128GB DGX Spark and put it on the accuracy versus task-time Pareto frontier, and NVIDIA reports LiLiCorr speculative decoding gives up to 2.3x faster generation there. The Nemotron 3.5 Super VL DGX Spark guide goes deeper.
It also plugs into the agent harnesses businesses already use: OpenCode, OpenClaw, Pi, LangChain, Cline, OpenHands and Hermes Agent.
📺 Watch: New FREE Nvidia Nemotron 3 Super + Ollama
The Speed and Scale Numbers
For a business, speed means cost. Faster tokens on the same hardware means cheaper agent runs. Here is what NVIDIA reports:
- 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.
- 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.
That last one is the line a CFO cares about. Serving 8x as many users on the same GPUs changes what an agent product costs to run.
Customer Results NVIDIA Cites
NVIDIA named several companies in its launch material. These are their figures as reported by NVIDIA, not independent tests.
- Scale AI (Skydance): 40x lower inference cost and 5x lower latency, per NVIDIA.
- Ramp: up to 50% higher throughput and 10x lower latency, which NVIDIA describes as expected results.
- AI21 and Alice: a 41% accuracy improvement, NVIDIA reports.
In my video I also mentioned one company's claim of 10,500 tokens per second across users on a single DGX Station. That is their own case study claim, so treat it that way.
On benchmarks, NVIDIA reports 63.44% on Hermes Bench versus 56.33% for Nemotron 3 Ultra, and 87.05% on LangChain evals versus 84.3% for Ultra. On CharXiv chart reasoning it scores 80.8%, or 84% with Python code interpretation. No one has independently replicated these yet. I test models on real business tasks over on Goldie Bench.
What It Means for Businesses Using Agents
Here is the practical takeaway.
Document and screenshot work just got cheap. Most business processes start with a file: an invoice, a contract, a report, a chart, a screenshot of a dashboard. A free, open model that can read all of those and then call tools covers a huge share of real agent work.
You can mix models. I use it free in OpenCode inside my Agentic Operating System. I asked it to "build a beautiful to-do list app, make it entertaining, engaging" and it built a working app. I also run it in Hermes Agent. But I am honest about it: it is not a frontier model and it is not going to replace Opus 5.5. The smart play is to route file-and-picture jobs to Nemotron and keep your frontier model for the hardest reasoning.
Private AI is realistic. With open weights, open data and a desk-sized option like DGX Spark, a company can keep sensitive documents in-house and still run capable agents.
Your costs can drop. If the customer numbers hold up anywhere close to what NVIDIA cites, running agents on an open model like this beats paying per call for everything.
My Agent OS runs Hermes, OpenClaw, OpenCode and Claude side by side, so dropping in a new open model takes seconds. The Agent OS guide shows how it is set up.
Want to map out where this fits in your own business? Book a free AI strategy session and we will work through it.
Or join the AI Profit Boardroom ($69/mo) to get the Agent OS zip, a video tutorial, a 30-day roadmap, updates, 4 weekly coaching calls and daily tutorials, with 3,400+ members building alongside you.
FAQ
When did NVIDIA release Nemotron 3.5 Super VL?
15 October 2026, with a technical blog, open weights, training recipes, data, RL environments, 40 vision-language datasets on Hugging Face and a GitHub cookbook.
Can businesses use it commercially?
It ships under the OpenMDW License Agreement v1.1 with open weights. Read the license terms for your specific use before you build a product on it.
What is the easiest way for a company to deploy it?
If you are already on a cloud, use Amazon SageMaker JumpStart, Microsoft Foundry, Baseten, Fireworks AI or Together AI. If you run NVIDIA infrastructure, use the NIM and Dynamo recipes. To test it first, try OpenRouter's free variant.
Are the customer results independently verified?
No. The Scale AI, Ramp, AI21 and Alice figures are reported by NVIDIA. EXO Labs' DGX Spark evaluation is the main outside check so far.
If you want help turning this launch into lower agent costs, grab a free AI strategy session. NVIDIA Nemotron 3.5 Super VL is one of the strongest open options a business can deploy today.
