Nemotron 3.5 Super VL: NVIDIA's open model that sees video, and the hardware it needs

Nemotron 3.5 Super VL is NVIDIA's open-weights model that reads text, images, documents, and video in a single system. It has 120 billion parameters, but only 12 billion are active per generated token, and NVIDIA's own table does not rank it first on coding benchmarks. Here is what it does, how much hardware it really needs, and whether running it yourself is worth it.

What changed with Nemotron 3.5 Super VL

Released on October 15, 2026, the model is not just another vision model. It is one model that reasons, uses tools, and understands text, images, documents, and video. Until now you had to stitch a language model to a vision model with glue code. NVIDIA distilled this one from its larger Nemotron 3 Ultra and released it open: weights, training recipes, reinforcement learning environments, and the datasets. The license is OpenMDW, and the model card says it is fine for commercial use, so read the full text before selling it as a service.

How it works inside: 12 billion active, 120 billion loaded

It is a mixture of experts. For each word it generates, only about 10 percent of the parameters do any work. The catch is that the full 100 percent has to sit in memory, because you cannot know in advance which experts a token will need. The model also mixes Mamba layers with attention, accepts up to a million tokens of context, and includes a multi-token prediction head to speed up generation. The difference between what it computes and what it occupies drives every hardware decision below.

The walkthrough: from download to first answer

1. Download NVIDIA's checkpoint from Hugging Face.
2. Start the server with vLLM. The model card recipe uses four H100 GPUs, tensor parallelism of four, and 85 percent GPU memory utilization, and it serves on port 8000.
3. Send a photo of a receipt and ask for JSON, or a video of up to two minutes.
4. Know where it gets stuck: the recipe allows one image and one video per request, so ten photos means splitting the job. For 1080p video the card recommends one frame per second with a cap of 128 frames, not the default two per second.

One more detail: reasoning and the answer share the same token budget. Setting enable thinking to false makes the model answer directly.

The hardware math

In BF16, 120 billion parameters at two bytes each come to 240 GB of weights alone. Four 80 GB H100s total 320 GB, and at 85 percent utilization that leaves about 32 GB for context and video after the weights. That is tight. At four bits, the weights drop to roughly 60 GB, which fits in a DGX Spark with 128 GB of unified memory, leaving about 68 GB before the system and cache. Reading roughly 6 GB of active weights per token at 273 GB per second puts the theoretical ceiling near 45 tokens per second. That figure is WAiBOT's own analysis, not a measurement.

Three use cases for Nemotron 3.5 Super VL

• Receipts to JSON. An accounting firm or small business can turn hundreds of receipt photos into vendor, date, amount, and taxes. Ramp, which does this with receipts, expects up to 50 percent better performance and ten times lower latency after post-training. That is their claim, in their workflow. What you can charge for is the service of keeping it running and checked, since the model itself is open.
• Charts to data. On CharXiv, a chart reasoning test, NVIDIA reports 80.8 percent, versus 63 percent for the previous model in the family and 84 percent when it can use Python. Give it the interpreter, not only the image. Running it yourself also keeps client documents on your own server.
• Short video review. Two minutes at one frame per second is 120 frames, inside the 128 the card suggests. This case is WAiBOT's analysis, not NVIDIA's, so you need to build your own tests with your own videos.

Who should skip it

Skip it if you want to run it on a regular laptop: even at four bits it needs about 60 GB of weights, and the card mentions Blackwell and Hopper GPUs on Linux. Skip it if you only want the best coding agent: in NVIDIA's own table, open models beat it on Terminal-Bench by almost 20 points. And skip it if you process ten documents a month, where a hosted service works out better than setting up a server.

The objection and the nuance

On GPQA, LiveCodeBench, SWE-bench, and Terminal-Bench, none of the four rows ranks Nemotron 3.5 Super VL first. But that table mostly measures reasoning and code, not cost per task, speed, or reading documents and video. NVIDIA is selling speed and adaptation to a specific task. The "up to five times faster" claim applies to agentic tasks, with a single GPU, against other open models of similar size, and the blog does not say which GPU or which models. Scale AI's expectation of costs forty times lower comes after post-training for one workflow. EXO Labs places the model on the accuracy and speed frontier on a DGX Spark, which helps, but it is one machine and one test.

What breaks in practice

Reasoning can spend the whole token budget and never answer if the cap is low. The recipe accepts one image and one video per request, only text comes out, and the BF16 checkpoint does not accept audio. Much of the post-training data is synthetic, and the blog reports no safety evaluations. Memory is the limit you hit first: a long video eats the 32 GB of headroom.

How to apply it in your business?

1. Weights equal parameters times bytes: 120 billion times two is 240 GB, and times half a byte, 60 GB.
2. Before buying hardware, try a hosted version on OpenRouter, Together, or Fireworks with 20 of your real documents.
3. Measure accuracy, time per document, and cost per document.
4. If the math works and your data cannot leave your company, move to the NVFP4 version on a DGX Spark.
5. Read the full OpenMDW license before selling it as a service.

Twelve billion active parameters does not mean twelve billion of memory.

Frequently asked questions

Is Nemotron 3.5 Super VL free for commercial use?

The model card says the OpenMDW license allows commercial use. Read the full license text before selling it as a service.

How much memory does Nemotron 3.5 Super VL need?

About 240 GB of weights in BF16, which the model card recipe runs on four 80 GB H100 GPUs, and about 60 GB at four bits, based on WAiBOT's estimate.

Is it really five times faster?

NVIDIA says up to five times on agentic tasks with a single GPU, against open models of similar size. The blog does not name the GPU or the models, so treat it as a hypothesis to measure yourself.

Can it process video?

Yes, clips of up to two minutes, with one frame per second recommended for 1080p and a cap of 128 frames, per the model card.

Conclusion

The question is not whether Nemotron 3.5 Super VL is the smartest model. According to its own card, it is not. The question is whether your bottleneck is reading documents and video at volume. If it is, test it with your own files first. Which document or video from your work would you try on it first?