Nemotron 3.5 Super VL

License OpenMDW 1.1

License OpenMDW 1.1

TL;DR

• 120B total and 12B active: a Mamba-2 and Transformer hybrid MoE, 512 routed experts, multi-token prediction head
• Beats size peer Qwen3.5-122B-A10B on 10 of 13 text and agent rows, loses to it on 7 of 9 vision rows
• OpenMDW 1.1 with no attribution or revenue clause, and the pretraining and post-training datasets published

☍ Announcement ⬇ Download Model

System Requirements

RAM	Data center; 4x H100 or 2x GB200
GPU	NVIDIA Blackwell or Hopper; NVFP4 build recommended for DGX Spark
VRAM	120B total with 12B active; BF16, NVFP4 and FP8 builds published

✓ Ollama

Table of Contents

1. What it is
2. Compare like with like
3. The VL half is the weak half
4. What is actually open here
5. Running it
6. The community got there first
7. Who should use it
8. Sources and further reading

Read NVIDIA's own benchmark table for Nemotron 3.5 Super VL and the model looks mediocre. It loses most rows. Then you notice the table compares a 120B model against a 763B one, a 550B one, and whatever Qwen and Z.ai are shipping this month, and that exactly one competitor in the list is the same size. Against that one, Nemotron 3.5 Super VL wins ten of thirteen text and agent benchmarks. It also loses seven of nine vision benchmarks, which is an odd thing for a model with VL in its name.

What it is

Released 12 October 2026: 120B total parameters, 12B active, a Mamba-2 and Transformer hybrid Mixture-of-Experts with 512 routed experts per MoE layer and a shared multi-token prediction head for speculative decoding. Vision comes from a C-RADIO encoder plus a vision-language connector. Context runs to 1M tokens. Inputs are text, images and video up to two minutes of MP4; output is text only.

It is the mid-tier sibling of Nemotron 3 Ultra, the 550B-A55B flagship we covered in June, and it inherits that model's two best properties: the hybrid Mamba architecture, which is what makes a million-token context affordable to serve, and the licence.

Three builds shipped at once, BF16, NVFP4 and a base checkpoint, which is more than most labs manage on day one.

Compare like with like

NVIDIA's published table sets Nemotron 3.5 Super VL against six competitors, and only one of them, Qwen3.5-122B-A10B, is in the same weight class. The rest are larger, newer-generation, or both. Here is the comparison that actually tells you something, with every figure taken from NVIDIA's model card.

Ten wins, three losses, and the wins are not narrow. Twenty points on BrowseComp and twelve on Terminal-Bench 2.1 against a model with two more billion total parameters is a real result that NVIDIA's own presentation buries by sitting it next to DeepSeek-V4-Flash and Nemotron 3 Ultra.

The telecom row is the one to take seriously as a loss. Fourteen points is not noise, and tau2-bench telecom is a tool-calling benchmark where almost everything else on NVIDIA's table scores in the high eighties or nineties. Something in that domain does not suit this model.

The VL half is the weak half

Now the part that should change how you deploy it. Against the same size peer, on vision:

Seven losses, two narrow wins, and three of those losses are inside a point. So it is not bad at vision; it is unremarkable at vision while being genuinely strong at text and agent work. The headline capability in the model's name is the one where it is most ordinary.

Video is weak across the entire field, not just here. 32.85 on Video-MME v2, with the best model on the table at 45.1, says the two-minute MP4 support is a feature to prototype with rather than build on.

Worth noting that NVIDIA flags several of these rows as run with internal scaffolding or official open-source implementations rather than their standard NeMo Gym pipeline, including BabyVision, CharXiv, MathVision, MMLongBench-Doc, OCR-Reasoning, OCRBench-v2, V* and Video-MME-v2. That is most of the vision table. Cross-model vision comparisons are the least trustworthy numbers on any card.

What is actually open here

This is the reason to care, and it is the same reason Nemotron 3 Ultra mattered. The licence is OpenMDW-1.1, the Linux Foundation's Open Model Definition for Weights. No attribution string in your UI, no revenue trigger, no acceptable-use exhibit, no separate agreement to negotiate when you grow. Commercial or non-commercial, both fine.

Set that against the month's other big open-weight release. Meta's Muse Spark Open requires you to display "Built with Muse Spark Open" in your interface, bans operating critical infrastructure without defining the term, and makes you sign a separate deal with Meta if your group clears $20 million in revenue while selling model access. Both are called open weights. They are not the same thing.

NVIDIA also publishes the data. The pretraining corpus is over 20 trillion tokens across English, nineteen other spoken languages and 43 programming languages, and the component datasets are on Hugging Face as real downloads, not as a bibliography: Nemotron-CC-v2, Nemotron-Post-Training-Dataset-v1, Nemotron-Math-v2, Nemotron-Terminal-Corpus and others, several with tens of thousands of downloads of their own.

Be precise about the limits of that, though. The card also lists private third-party datasets and private NVIDIA datasets that are not publicly accessible, and much of the post-training data is synthetic, generated with DeepSeek-V3.2, Qwen3.5, GPT-OSS-120B and Kimi-K2. So this is not a fully reproducible model. It is substantially more open about its inputs than any frontier lab, which is a different and still valuable claim.

Running it

Four H100s or two GB200s, Linux, Blackwell or Hopper. vLLM and SGLang both have recipes at TP4 or TP2, Transformers works with trust_remote_code=True, and Docker Model Runner is supported.

vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 \
  --tensor-parallel-size 4 --trust-remote-code

NVIDIA recommends the NVFP4 build for DGX Spark and data-center GPUs, which puts a 120B multimodal model with a long context inside reach of a desk-side box. If you have been following our dual DGX Spark work, that is the configuration to try first.

Two deployment notes from the card that will bite you otherwise. Reasoning tokens and the final answer share the same max_tokens budget, so size it generously. And the published deployment recipes may use a context window smaller than the advertised 1M, so check the recipe rather than the headline. For 1080p video, NVIDIA suggests 1 FPS and up to 128 frames.

The community got there first

One detail worth recording. The ggml-org GGUF conversion has 29,351 downloads against 15,095 for NVIDIA's own BF16 release, and Unsloth's GGUF has another 11,677. Within three days of launch, a community quantization was the most-downloaded way to get this model.

Domain fine-tunes arrived almost as fast: Fastino published finance and healthcare variants on 14 October, two days after release. A permissive licence plus published training data produces that; a gated download and a revenue clause does not.

Who should use it

Deploy it if you need an agentic model you can serve on four H100s with no licence conditions attached. BrowseComp at 78, tau2-bench airline at 76.5 and GraphWalks at 89.71 are the rows to build on, and the Mamba hybrid is what keeps long contexts affordable in production rather than just possible.

Fine-tune it if you have a domain. The base checkpoint shipped alongside the instruct model, the post-training datasets are public, and two useful verticals already exist to copy from.

Look elsewhere if vision is your primary workload. On its own card it trails its size peer on seven of nine vision rows, and the frontier open models are further ahead still. For document and retrieval work specifically, pairing a smaller specialist with a proper embedding model will beat leaning on this model's VL half.

And if you are choosing between open-weight releases this month, read the licences before the benchmarks. A model that loses a few points and lets you ship without a lawyer is usually the better engineering decision, and that is the trade NVIDIA keeps making on purpose.

Sources and further reading

• Nemotron 3.5 Super VL BF16 model card and full benchmark tables
• The NVFP4 build, recommended for DGX Spark
• Nemotron post-training dataset
• Nemotron-CC-v2 pretraining corpus
• The OpenMDW 1.1 licence text

Your under-ten-minutes move: open the model card, find the Qwen3.5-122B-A10B column, and read only that column against the Nemotron one. Ignore everything larger. That is the comparison that tells you whether this model is good, and it is the one the table is laid out to obscure.

Tested on: not independently tested. All benchmark figures, parameter counts, dataset lists and hardware requirements are NVIDIA-reported and were read from the Hugging Face model card on the date below. NVIDIA states that several vision rows were produced with internal scaffolding rather than its standard evaluation pipeline. Download counts were read from the Hugging Face API on the same date.
Date checked: 2026-10-19