For agent work, Nemotron 3.5 Super VL beats Nemotron 3 Ultra on NVIDIA's own numbers, and it's the smaller, faster model of the two. That's the short answer to Nemotron 3.5 Super VL vs Nemotron 3 Ultra. The new model scores higher on Hermes Bench and on LangChain evals, it can see images and documents, and it was built by learning from Ultra in the first place.
📺 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
That last part is what makes this interesting. Here the student beats the teacher on the agent tests NVIDIA chose to publish. Below: the numbers, why the smaller one wins, where Ultra still fits, and how to check it on your own work.
If you want to see how I'm plugging new open models like this into a working setup, that's what we do every week inside the AI Profit Boardroom.
The quick verdict
If you're building agents, pick Super VL first. Per NVIDIA, it wins both agent benchmarks they compared against Ultra, and it runs with only 12B parameters active per token. That means it's cheaper to serve and quicker to respond.
Ultra isn't dead. It's the model Super VL was distilled from, so it's still a strong text model. But NVIDIA doesn't publish a full side-by-side across every category, so if you have a job where Ultra already works well for you, test before you switch.
For a full rundown of the new model on its own, read my Nemotron 3.5 Super VL overview.
Nemotron 3.5 Super VL vs Nemotron 3 Ultra: head-to-head
Here's everything NVIDIA has actually said that lets you compare the two directly. I've left out anything they didn't publish, so you won't find a made-up parameter count for Ultra here.
| Category | Nemotron 3.5 Super VL | Nemotron 3 Ultra |
|---|---|---|
| Hermes Bench (per NVIDIA) | 63.44% | 56.33% |
| LangChain evals (per NVIDIA) | 87.05% | 84.3% |
| Relationship | Distilled from Nemotron 3 Ultra | The teacher model Super VL learned from |
| Size | 120B total, 12B active per token (mixture-of-experts) | Not stated in this comparison |
| Inputs | Multimodal: text, images, documents, video frames | Not stated as multimodal in this comparison |
Two benchmarks, two wins. Hermes Bench is the bigger gap: just over seven points. LangChain evals is closer, under three points, but still a win.
Want the wider benchmark picture, including chart reading and OCR? I break it all down in the Nemotron 3.5 Super VL benchmark page.
📺 Watch: NEW Nemotron 3 Ultra is Insane (FREE!) 🤯
Why the smaller model wins
Most people assume more parameters equals better results. That's the old way of thinking about models. The new way is that how a model is trained and how it generates matter just as much as raw size. Two things explain the gap here.
1. Distillation from Ultra
Super VL was distilled from Nemotron 3 Ultra. In plain terms, Ultra acted as the teacher and Super VL learned from its outputs. The student keeps a lot of what the teacher knows, then NVIDIA trains it further for agent work: reasoning, instruction following and tool use, all in one model.
2. Multi-token prediction (MTP)
Super VL uses multi-token prediction. An internal drafter proposes several tokens at once, and the model checks them in a single pass. When the guesses are right, you get more tokens out per step.
That matters a lot for agents. An agent doesn't answer once. It thinks, calls a tool, reads the result, thinks again, and repeats. Faster generation across dozens of steps adds up.
3. Only 12B active per token
Super VL has 120B parameters in total but only 12B are active for each token, thanks to its mixture-of-experts design. It also uses a hybrid architecture that mixes Mamba layers with attention. The result is a model with a big brain that only uses the part it needs at each moment.
NVIDIA reports up to 5x higher single-GPU throughput on agentic tasks versus open models its size. That's a comparison against other open models, not against Ultra, but it tells you where the design is pointed: speed on agent loops.
4. It can see
Super VL reads text, images, documents and video (through sampled frames or short clips). It outputs text. For agent work that starts with a screenshot, a PDF, a chart or a receipt, that's a real advantage. Your agent can look at the thing instead of you describing it.
When I tested it inside OpenCode, I just picked Nemotron 3.5 Super VL in the model drop-down and asked it to build a beautiful to-do list app and make it entertaining and engaging. The build came out working. Not frontier-level, not replacing Opus 5.5, but free, fast, and good enough for a lot of everyday agent jobs.
If you'd like help mapping which of your jobs suit a model like this, book a free AI strategy session.
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
When Ultra might still make sense
Here's the honest bit. NVIDIA published two agent benchmarks against Ultra. That's it. They don't publish a full comparison across every type of task, so there's no table that proves Super VL is better at everything.
That leaves a few cases where Ultra could still be the right call:
- You already have Ultra running and tuned. If your prompts, pipelines and evals are built around it and the results are good, switching has a cost. Test first.
- Your job isn't agent work. The published wins are on agent and tool-use benchmarks. If you're doing long-form text tasks that don't involve tools, you don't have NVIDIA numbers either way.
- You have a specific edge case. Some prompts just work better on one model. You only find that out by running your own jobs on both.
The simple rule: if you're starting fresh on an agent project, start with Super VL. If you're already on Ultra and happy, run a side-by-side before you move.
How to test this on your own jobs
Benchmarks give you a starting point. Your own work gives you the answer. Here's how I'd run the check:
- Pick five to ten real tasks you actually run every week. Not demo prompts. Real ones.
- Run each task on both models with the same prompt and the same tools.
- Score them on three things: did it finish, was the output right, and how long did it take.
- Add at least one task that starts with an image or a PDF.
- Keep whichever model wins most of your jobs, and keep the other one as a fallback.
I use the same thinking behind Goldie Bench: real tasks, scored plainly, so you can tell what's actually useful versus what just looks good on a chart.
To try Super VL without a big GPU, use it hosted. It's on OpenRouter (including a free variant) and you can pick it free inside OpenCode. If you want it local, the Ollama name is nemotron-3.5-super, and you can run it with: ollama run nemotron-3.5-super. Just know it needs roughly 70 to 90GB of memory, so it's not a normal laptop model. The Nemotron 3.5 Super page covers that Ollama listing in more detail.
The caveat: these are NVIDIA's numbers
Every benchmark on this page comes from NVIDIA. Hermes Bench 63.44% vs 56.33%, LangChain 87.05% vs 84.3%: all NVIDIA-reported. There's no independent replication of this head-to-head yet.
That doesn't mean the numbers are wrong. It means NVIDIA picked the tests, and companies tend to show the comparisons they win. So treat the result as a strong signal, not a final word, and back it up with your own testing.
The good news is that Super VL is open weights under the OpenMDW License, and it's free to try hosted. Checking it yourself costs you very little.
Where this fits in your setup
I don't pick one winner for everything. Inside my Agent OS, I run Hermes, OpenClaw, OpenCode and Claude side by side, and I swap in open models like Super VL when a job suits them.
For me, Super VL slots in for jobs that start with a picture or a file: reading screenshots, pulling data out of PDFs, checking charts. If I run out of tokens in OpenCode, I switch to another free model there or use Super VL inside Hermes Agent, where it shows up as one of the models I can pick.
If you want the full setup, the Agent OS guide walks through how the pieces connect. And if you want the Agent OS zip, the video tutorial, the 30-day roadmap and four weekly coaching calls, join the AI Profit Boardroom for $69/mo. There are 3,400+ members in there already building with this stuff.
FAQ
Is Nemotron 3.5 Super VL better than Nemotron 3 Ultra?
For agent work, yes, on NVIDIA's numbers. It scores 63.44% on Hermes Bench vs 56.33% for Ultra, and 87.05% on LangChain evals vs 84.3%. NVIDIA doesn't publish a full comparison across every task type, so test your own jobs.
How can a smaller model beat the one it was distilled from?
Super VL learned from Ultra through distillation, then was trained further for reasoning, instruction following and tool use. It also uses multi-token prediction and a mixture-of-experts design with only 12B active parameters per token, which makes it fast on long agent loops.
Can Nemotron 3 Ultra read images like Super VL?
Super VL is multimodal: it reads text, images, documents and video frames. NVIDIA's comparison doesn't present Ultra as a multimodal model, so if your agent work starts with a screenshot or a PDF, Super VL is the one built for it.
Should I switch from Ultra right now?
If you're starting a new agent project, start with Super VL. If Ultra already works well for you, run five to ten of your real tasks on both first. If you want a second pair of eyes on that plan, grab a free AI strategy session.
