Nemotron 3.5 Super VL: NVIDIA's Open 120B Vision Model

Nemotron 3.5 Super VL is NVIDIA's new open vision-language model, and it was one of the headline items in a batch of open-model news NVIDIA published on October 21. It reads text, images and short videos, keeps a context of up to 1 million tokens, and runs with only 12 billion of its 120 billion parameters active at a time. The weights are on Hugging Face under NVIDIA's open model license.

What actually happened

NVIDIA's October 21 blog post, by Kari Briski, rolled out a list of open models and tools aimed at European industry and developers. Alongside the new vision-language model, it mentioned Cosmos 3.5 for physical AI, Alpamayo 2 Edge for autonomous driving, a protein design model in BioNeMo that NVIDIA says is 13 times faster than RFdiffusion3, and partners including Mistral, Black Forest Labs and H Company. The post says the models are available on Hugging Face, ModelScope, OpenRouter and as NIM microservices on build.nvidia.com.

The Hugging Face model card fills in the details. The full name is NVIDIA-Nemotron-3.5-Super-VL-120B-A12B. It uses a C-RADIO vision encoder, accepts video files up to about two minutes, supports multi-token prediction for faster decoding, and is sized to run on four H100 GPUs or two GB200 chips. The license is the OpenMDW 1.1 open model license. If you followed our NVIDIA Kumo tabular model story, this is the same open-model push, now aimed at multimodal agents.

The numbers, with the caveats

All of these come from NVIDIA's model card, run by NVIDIA:

• GPQA-Diamond: 89.52.
• Humanity's Last Exam, text only: 31.79.
• SWE-bench Verified: 71.10.
• LiveCodeBench v6: 88.22.
• BrowseComp: 78.
• OSWorld-Verified: 51.14.

The card is fair enough to show where it trails. On MMLU-ProX, a multilingual test, it scored 76.91, below the 83.2 listed for Qwen3.5-122B. Its AA-Omniscience hallucination score was 30.0, which shows it still makes things up. No independent lab had published a full replication at the time of writing.

Open weights plus a 1 million token window means you can run long agent sessions on your own hardware, if you own a rack of H100s.

Why Nemotron 3.5 Super VL matters

Most strong open models are text only, and most vision models struggle with long agent sessions. This one tries to do both: see a screen or a chart, read a long document, and act. The OSWorld score, a test of using a real computer, points straight at agent use cases.

The 12B active parameter count is the real trick. You need memory for all 120B weights, but each token only uses a slice, which keeps responses fast. That is why it fits on a single four-GPU node rather than a cluster.

For the open-model race, it gives builders another non-Chinese option at the high end, next to Mistral, which we covered in our Mistral Voxtral 3 piece. Europe-focused partners and the OpenMDW license make it easier for companies that need to host models themselves.

What is still unclear

NVIDIA did not publish per-token prices for its hosted endpoint in the posts we read, and third-party hosts will set their own rates. Real-world speed on four H100s will depend on context length, since a full 1 million token prompt is very different from a short chat. Video input is limited to short clips, so it is not a tool for hours of footage. And the Hugging Face card is dated October 15, a week before the blog post, so early testers may already have numbers worth checking.

What to do as a reader

If you build agents, try Nemotron 3.5 Super VL through OpenRouter or NVIDIA's hosted endpoint before buying hardware. Test it on your own screenshots, charts and long files, not just on benchmark-style questions, and check the hallucination rate on your data.

If you trade AI narratives, remember that open model launches rarely move NVIDIA's stock and have nothing to do with crypto tokens using AI branding. Nothing here is investment advice. Wait for Friday's Bolt release if you need something small enough for a laptop.

Sources: NVIDIA blog on open models and Hugging Face model card.