In Nemotron 3.5 Super VL vs Nemotron 3 Ultra, the smaller, faster model wins where it counts for agents: it beat Nvidia’s bigger flagship on Hermes Bench, LangChain’s agent tests and several others, and it can see. I cover the head-to-head in the video below. Here’s exactly where each one wins and which to use.
Key takeaways
- Super VL beats Ultra on Hermes Bench (63.44% vs 56.33%) and LangChain (87.05% vs 84.3%).
- Super VL adds vision and is far stronger on documents and OCR.
- Ultra still edges it on some tests, like tau2-bench and MMLU-ProX.
- Super VL is faster and cheaper to run.
Head-to-head: agent tests
| Agentic benchmark | Nemotron 3.5 Super VL | Nemotron 3 Ultra |
|---|---|---|
| Hermes Bench (Nvidia-reported) | 63.44% | 56.33% |
| LangChain behavioural evals (Nvidia-reported) | 87.05% | 84.3% |
| BrowseComp | 78 | 67.85 |
| Terminal-Bench 2.1 | 60.67 | 54.80 |
| SWE-bench Pro | 51.78 | 49.66 |
| OSWorld-Verified | 51.14 | N/A |
Head-to-head: vision and documents
| Vision benchmark | Nemotron 3.5 Super VL | Nemotron 3 Ultra |
|---|---|---|
| CharXiv reasoning (charts) | 80.82 | 77.8 |
| OCRBench v2 (English) | 68.12 | 54.2 |
| MMLongBench-Doc (long documents) | 62.82 | 49.06 |
| V* (visual search) | 85.60 | 70.5 |
| OCR-Reasoning | 65.08 | 54.7 |
| MathVision | 78.16 | 78.8 |
This is where Super VL pulls clearly ahead — it was built as a vision-language model, while Ultra is primarily a text reasoning model.
Want this working in your business, not just bookmarked? Using free open models in your AI agents is exactly what we build together inside the AI Profit Boardroom — 3,400+ members, four live calls a week, and a full one-hour DeepSeek Harness course in the classroom.
Prefer it mapped 1-on-1 first? Book a free strategy session and we’ll plan it for your exact situation.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →
Where Nemotron 3 Ultra still wins
| Benchmark | Nemotron 3.5 Super VL | Nemotron 3 Ultra |
|---|---|---|
| MMLU-ProX | 76.91 | 83.1 |
| IFBench (instruction following) | 70.86 | 82.9 |
| tau2-bench telecom | 73.19 | 83.0 |
| AA-LCR v1.1 (long-context reasoning) | 68.62 | 75.2 |
| SWE-bench Verified | 71.10 | 73.9 |
So Ultra remains strong for some knowledge and instruction-following tasks. But for most agent workflows, Super VL’s wins matter more.
Why the smaller model wins
Nvidia trained Super VL by learning from Ultra, and its design is built for speed: only 12B of its 120B parameters are active per token, and multi-token prediction lets it draft several words at once. Nvidia says it’s comparable with its largest reasoning model at roughly a fifth of the parameters — and it’s much cheaper to serve.
Which should you use?
| If you need… | Use |
|---|---|
| An agent model that reads screenshots and PDFs | Nemotron 3.5 Super VL |
| Speed and lower cost | Nemotron 3.5 Super VL |
| Strict instruction following on text | Nemotron 3 Ultra |
| Broad knowledge tests | Nemotron 3 Ultra |
Let me be clear: this is not a frontier model. It’s not going to replace Claude Opus 5.5 for your hardest work. What it is: a free, open, fast model that can read screenshots, PDFs and charts and then act on them — which is exactly the job a lot of agent work starts with.
What this means if you use Nemotron 3 Ultra today
If you already run Nemotron 3 Ultra in your agents, try Super VL as a drop-in for agent and document work. You’ll likely get equal or better results on tool use and browsing, plus the ability to read images, at lower cost and higher speed. Keep Ultra for the text-heavy jobs where it still leads, like strict instruction following, and compare both on your own tasks before switching everything.
The bigger lesson
Smaller, newer models beating bigger, older ones is now normal in AI. Better training data, learning from larger models and smarter architectures mean the size of a model is no longer the best guide to how good it is. For agent builders, that’s great news: faster, cheaper models keep getting good enough for more of the work.
Running costs compared
Super VL activates about 12B parameters per token, which makes it much cheaper to serve than a larger flagship. On OpenRouter it costs $0.15 per million input tokens and $0.50 per million output tokens, and there’s a free variant. For high-volume agent work, that difference in running cost adds up quickly.
Final thoughts
The short version: for agents, use Super VL by default. It scores higher on the agent tests that matter most, it can read screenshots and documents, and it’s cheaper and faster to run. Keep Ultra in mind only for the specific text-heavy tasks where it still leads, and test both on your own work before you commit.
Nvidia’s own framing says it all: Super VL is comparable with its largest reasoning model at roughly one-fifth the parameter count.
Whichever you choose, keep the other one a click away in a separate profile.
How I’d test it this week
The simplest way to decide is to run both on five of your own tasks — three agent jobs and two visual ones — and compare. Most people will find Super VL wins on the agent and visual work while being faster, and that settles it.
If Ultra wins on a particular text task you rely on, keep it for that job and use Super VL for everything else.
Common questions
Should I delete Nemotron 3 Ultra from my setup? Not necessarily — keep it if it handles specific text tasks well for you. Is Super VL just a smaller Ultra? No — it’s a different, newer model with vision built in, trained partly by learning from Ultra. Which is free? Super VL is free in OpenCode and has a free OpenRouter variant at the time of writing. Which is better for Hermes Agent? On Hermes Bench, Super VL scored higher, 63.44% vs 56.33%.
The bottom line on Nemotron 3.5 Super VL vs Nemotron 3 Ultra
For agents, Nemotron 3.5 Super VL beats Nemotron 3 Ultra: it scores higher on agent tests, it can see, and it’s faster and cheaper. Keep Ultra in mind for text-heavy instruction following.
FAQ: nemotron 3.5 super vl vs nemotron 3 ultra
Is Nemotron 3.5 Super VL better than Nemotron 3 Ultra?
For agent tasks and vision, yes — it scores higher on Hermes Bench and LangChain’s agent tests.
Why is the smaller model better?
It learned from Ultra during training, and it’s designed for speed and agent work.
Where does Ultra still win?
On some knowledge and instruction-following tests, like MMLU-ProX and IFBench.
Does Nemotron 3 Ultra have vision?
Super VL is the vision-language model; it’s much stronger on documents and OCR.
Which is cheaper to run?
Super VL — only 12B parameters are active per token.
Which model should I use in OpenCode?
Nemotron 3.5 Super VL is free in OpenCode at the time of writing and can read screenshots, which makes it the better everyday choice there.
Two ways I can help from here. Join the AI Profit Boardroom — it’s where your AI agent model setup gets built with 3,400+ members doing the same.
Or grab a free strategy session and bring your questions.
Join the AI Profit Boardroom →Claim Your Free SEO Strategy Session →
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency, best-selling author, and an AI educator with 428K+ YouTube subscribers. I’ve spent 10+ years in SEO and online business and I test every AI agent on my own work before I recommend it. Join the AI Profit Boardroom for the daily builds, or book a free strategy session to talk through yours.
Related reading
Last updated October 2026. This page is a living guide to nemotron 3.5 super vl vs nemotron 3 ultra — the facts here move fast and it is updated as they do.