Nemotron 3.5 Super VL
NVIDIA's open-weight vision-language model with 120B parameters and 1M-token context
Nemotron 3.5 Super VL is a vision-language model from NVIDIA Corporation, released on Hugging Face on 15 October 2026. It has 120 billion total parameters with 12 billion active, handles text, images and video, and is aimed at reasoning, coding, tool use and document understanding.
- 120B total parameters, 12B active per token
- Context window of up to 1M tokens
- Mamba-2, mixture-of-experts and attention layers with a C-RADIO vision encoder
- Multi-Token Prediction gives native speculative decoding
- Released on Hugging Face under OpenMDW-1.1 on 15 October 2026
What it is
Nemotron 3.5 Super VL is a vision-language model built by NVIDIA Corporation and published on Hugging Face on 15 October 2026. It carries 120 billion parameters in total, of which 12 billion are active for any given token, and it accepts text, images and video while producing text. NVIDIA positions it for reasoning, coding, tool use, document understanding and video understanding. The company describes it as aimed at developers who want to customise a post-trained model, adapt it to domain data, build quantised versions, or run full-precision inference and evaluation. It is offered under the OpenMDW License Agreement version 1.1 and is described as ready for commercial or non-commercial use.
How it came about
NVIDIA Nemotron is described by the company as a family of open models that ship with open weights, training data and recipes, intended for building specialised AI agents. The Super VL variant appeared on Hugging Face on 15 October 2026 as the full-precision BF16 release, alongside an NVFP4 version aimed at efficient inference on DGX Spark or data centre GPUs. A technical report covering Nemotron 3.5 Super and Lightning was published by NVIDIA Research. Evaluation results for the model and its baselines were gathered through NeMo Gym, with published recipes covering containers, prompts, inference parameters and scoring.
Inside the model
The model interleaves Mamba-2 layers, mixture-of-experts layers and attention layers. Mamba-2 is a state-space design that mixes in cheaper sequence processing alongside standard attention, while the mixture-of-experts setup routes each token through only part of the network, which is why 12 billion of the 120 billion parameters are active at once. Vision input is handled by a C-RADIO encoder feeding a vision-language connector. Context extends to 1 million tokens. Multi-Token Prediction enables native speculative decoding, where extra tokens are drafted ahead and verified. Reasoning can be switched on or off through the chat template, with a medium setting that generates fewer tokens at some cost to accuracy.
What it can do
NVIDIA reports results across agentic coding, reasoning, knowledge, multimodal, instruction following, long-context and multilingual tests, all collected via NeMo Gym. As published, the model scores 89.52 on GPQA-Diamond, 88.22 on LiveCodeBench v6, 71.10 on SWE-bench Verified, 85.60 on V* and 80.82 on CharXiv reasoning. Its AA-Omniscience hallucination rate is reported as 30.0. Some figures sit below competing models in the same table, for example MMLU-ProX at 76.91 and Video-MME v2 at 32.85. Several benchmarks, including BabyVision, CharXiv-RQ, MathVision and OCRBench-v2, were run with official implementations or internal scaffolding rather than the open tooling.
How to use it
Weights are downloadable from Hugging Face under the OpenMDW License Agreement version 1.1, which the company says permits commercial and non-commercial use. The BF16 release is the full-precision reference intended for customisation, domain adaptation and evaluation; an NVFP4 release targets efficient deployment. Serving examples use vLLM or SGLang, with the API model name nemotron-3.5-super-vl. NVIDIA lists deployment on 4× H100 or 2× GB200 hardware.
Questions people ask about Nemotron 3.5 Super VL
What kind of model is Nemotron 3.5 Super VL?
It is a vision-language model, meaning it reads text, images and video and answers in text. NVIDIA built it for reasoning, coding, tool use, document understanding and video understanding, and released the weights openly on Hugging Face in October 2026.
How large is it and how much runs at once?
The model holds 120 billion parameters in total, but only 12 billion are active for each token, because its mixture-of-experts layers route work through a subset of the network. That keeps inference cheaper than the raw parameter count suggests.
What licence does it use?
The weights are released under the OpenMDW License Agreement, version 1.1. NVIDIA states the model is ready for both commercial and non-commercial use, and the BF16 files are meant for customisation, domain adaptation and full-precision evaluation.
What hardware do I need to run it?
NVIDIA lists 4× H100 or 2× GB200 for deployment. The BF16 card shows serving through vLLM or SGLang, and a separate NVFP4 release exists for efficient inference on DGX Spark or data centre GPUs.
Can I turn its reasoning on and off?
Yes. Reasoning is configured through the chat template with enable_thinking set to true or false, and there is a medium level that produces fewer tokens than full reasoning, with a possible drop in accuracy.
Good reading and sources
There is no English Wikipedia article on Nemotron 3.5 Super VL yet, so the facts on this page come from the sources listed above: huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.