The Nemotron 3.5 Super VL benchmark results tell a clear story: it’s strongest on agent tasks and vision, beats Nvidia’s bigger Nemotron 3 Ultra on several agent tests, and trails the very best open models like Qwen3.8-Flash-Next on others. I walk through the headline numbers in the video below. Here’s every score that matters, in plain English.
Key takeaways
- Hermes Bench: 63.44% vs 56.33% for Nemotron 3 Ultra (Nvidia-reported).
- CharXiv (charts): 80.82; OCRBench v2 English: 68.12.
- SWE-bench Verified: 71.10; BrowseComp: 78.
- Lowest hallucination rate among the comparison models except Ultra.
Agent benchmarks
| 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 |
These are the numbers that matter most for agents like Hermes. Super VL beats the bigger Ultra on Hermes Bench, LangChain’s behavioural tests, BrowseComp, Terminal-Bench and SWE-bench Pro. Nvidia credits learning from Ultra during training.
Vision benchmarks
| 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 |
CharXiv tests whether an AI really understands charts; OCRBench tests reading text in images; MMLongBench-Doc tests long documents. Super VL is a big step up from Ultra on documents and OCR.
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 →
How it compares with Qwen
| Benchmark | Nemotron 3.5 Super VL | Qwen3.8-Flash-Next | Qwen3.5-122B-A10B |
|---|---|---|---|
| GPQA-Diamond | 89.52 | 92.7 | 86.7 |
| SWE-bench Verified | 71.10 | 78.7 | 71.9 |
| Terminal-Bench 2.1 | 60.67 | 85.80 | 48.70 |
| BrowseComp | 78 | 89.0 | 58.1 |
| CharXiv reasoning | 80.82 | 85.8 | 83.5 |
| OCRBench v2 (English) | 68.12 | 67.9 | 68.3 |
| GraphWalks (long context) | 89.71 | 76.1 | 86.1 |
| AA-Omniscience hallucination rate (lower is better) | 30.0 | 41.3 | 84.5 |
Qwen3.8-Flash-Next beats it on most tests, but Super VL wins on long context (GraphWalks) and has a much lower hallucination rate than the Qwen models listed. More in my Nemotron 3.5 Super VL vs Qwen comparison.
Speed claims
Nvidia says Super VL is five times faster than the next-best open agentic model in its class, and that its Dynamo serving software lets organisations serve up to eight times more users with 85% lower latency. These are Nvidia’s own figures — independent replication wasn’t available at launch, so treat them as vendor claims.
What the benchmarks mean for you
- Agent work: a strong free choice, especially for tool use and browsing.
- Documents and charts: one of its best areas.
- Coding: solid, but use a top model for hard production code.
- Accuracy: a lower hallucination rate is valuable for agents acting on their own.
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.
How to read these benchmarks
Benchmarks measure narrow skills under fixed conditions, so use them to shortlist models, not to crown a winner. Agent benchmarks like Hermes Bench, BrowseComp and Terminal-Bench are closest to how agents actually work. Vision tests like CharXiv and OCRBench tell you how well a model reads charts and text in images. Coding tests like SWE-bench reflect real software tasks. Always check whether a number is reported by the model maker or measured independently.
The scores I’d pay most attention to
For agent builders, three numbers stand out. Hermes Bench at 63.44% shows it handles agent workflows well for its size. A 30% hallucination rate on AA-Omniscience — far lower than several rivals — matters when an agent acts without supervision. And strong document and OCR scores mean it can genuinely read the screenshots and PDFs that so much business work starts with.
Test it on your own work
The best benchmark is your own: run the same three real tasks on Nemotron 3.5 Super VL and the model you use today — one visual, one research, one coding — and compare the results, speed and cost. That tells you more than any table.
Final thoughts
It’s also worth noting what the benchmarks don’t show. They don’t capture how easy a model is to access, how fast it responds in your region, or how well it follows your particular style. Nemotron 3.5 Super VL’s practical advantages — free access, open weights, a 1M-token context and vision in one model — don’t appear in any single score, but they matter a lot for everyday agent work.
Use the numbers above to decide whether it’s worth testing, then let your own results decide whether it stays.
How I’d test it this week
Turn the benchmarks into a quick test of your own: one chart, one long document, one agent task with tools and one coding task. Score each model you’re considering on accuracy, speed and cost. It takes an afternoon and saves weeks of guessing.
Repeat the same test whenever a new model launches, and you’ll always know which model deserves each job.
Benchmarks vs real-world results
In my own test, I asked it in OpenCode to build an entertaining to-do list app, and it delivered a working app quickly — consistent with its solid coding scores. For visual jobs, the strong document and OCR results match what you’d expect from a model built as vision-language from the start. Where it’s weaker on paper, like the hardest coding benchmarks, a frontier model is still the better choice. That’s exactly how I use it: the free, fast eyes of my agent stack.
The bottom line on the Nemotron 3.5 Super VL benchmark
The Nemotron 3.5 Super VL benchmark results show a fast, free model that punches above its size on agent and vision tasks. Test it on your own work — benchmarks are a starting point, not a verdict.
FAQ: nemotron 3.5 super vl benchmark
What did Nemotron 3.5 Super VL score on Hermes Bench?
63.44%, versus 56.33% for Nemotron 3 Ultra (Nvidia-reported).
Is Nemotron 3.5 Super VL good at vision?
Yes — 80.82 on CharXiv reasoning and 68.12 on OCRBench v2 English.
Is it better than Qwen?
Qwen3.8-Flash-Next scores higher on most tests, but Super VL wins on long context and hallucination rate.
How fast is it?
Nvidia says it’s five times faster than the next-best open agentic model in its class.
Where do these numbers come from?
Nvidia’s model card and launch materials.
Are the Hermes Bench numbers independent?
No — the Hermes Bench and LangChain figures are Nvidia-reported; the other scores come from Nvidia’s model card. Treat them as a guide and test on your own tasks.
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 benchmark — the facts here move fast and it is updated as they do.
