The Nemotron 3.5 Super VL Hugging Face release gives you everything: open weights in BF16 and NVFP4, an OpenMDW-1.1 licence that allows commercial use, and recipes for vLLM, SGLang and Transformers. Nvidia also released training data and recipes. My video below covers the model. Here’s what’s on Hugging Face and how to run it.
Key takeaways
- Checkpoints: BF16 (full precision) and NVFP4 (optimised).
- Licence: OpenMDW-1.1 — commercial use allowed.
- Hardware for BF16: 4× H100 or 2× GB200.
- Community GGUF builds (e.g. Unsloth) for llama.cpp-based tools.
What’s on Hugging Face
| Repository | What it is |
|---|---|
| nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 | Full-precision weights |
| nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 | Optimised 4-bit weights for Nvidia hardware |
| unsloth/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-GGUF | Community GGUF builds |
| Nvidia datasets | Open datasets for agentic tool use, long context, vision-language and more |
Hardware requirements
The model card lists 4× H100 or 2× GB200 GPUs on Linux (Hopper or Blackwell) for the BF16 release. The NVFP4 version is designed for more efficient deployment, including on Nvidia’s DGX Spark desktop machine. For consumer setups, use the GGUF builds in LM Studio or Ollama (70GB+).
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 to run it with vLLM
The model card’s vLLM recipe serves it with tensor parallelism across four GPUs, multi-token-prediction speculative decoding, and the served name nemotron-3.5-super-vl on port 8000, using the vLLM OpenAI image (nightly or v0.32.0). Tool calling uses the qwen3_coder parser. SGLang (v0.5.20) is supported on two GB200s, and Transformers works through the image-text-to-text pipeline with trust_remote_code.
Reasoning and video settings
- Reasoning modes: off, medium or on, set with
enable_thinkingand medium effort. - Reasoning and the answer share the token budget.
- Video: up to 256 frames at 2 FPS (1 FPS and up to 128 frames suggested for 1080p).
- Inputs: text, JPEG/PNG images, document pages and MP4 video up to 2 minutes.
The licence
Nemotron 3.5 Super VL is released under OpenMDW-1.1, and the model card says it’s ready for commercial or non-commercial use. As always, read the licence yourself before building a product on it.
Should you self-host it?
Self-hosting makes sense for companies that need privacy, control or fine-tuning — Nvidia released the recipes to make that possible. For everyone else, OpenRouter’s free and cheap paid versions are far simpler.
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.
Fine-tuning Nemotron 3.5 Super VL
Because Nvidia released training recipes and datasets, you can fine-tune the model on your own data — your documents, your formats, your way of working. That’s most useful for companies with a specific, repeated visual task, like reading a particular kind of invoice or form. Fine-tuning a 120B model needs serious GPUs, so most teams will use cloud GPU providers rather than their own hardware.
BF16 vs NVFP4
BF16 is the full-precision release: the best quality, but it needs the most memory. NVFP4 is a 4-bit format designed for Nvidia’s newer GPUs, cutting memory needs dramatically with little quality loss, which is what makes running it on smaller systems like DGX Spark possible. If you’re deploying on Nvidia Blackwell hardware, NVFP4 is usually the practical choice.
Hosted alternatives
If self-hosting isn’t worth it for you, the same model is available through OpenRouter, Nvidia NIM, ModelScope and other hosts, and Together AI added it at launch. That gives you the open model without managing GPUs.
Final thoughts
A last practical note: model cards are updated as recipes and benchmarks improve, so check the Hugging Face page for the latest serving instructions before you deploy. Nvidia’s recipes are tuned for its own hardware, and the community often publishes faster or smaller variants within days of a release — worth watching if you want to run it on less hardware.
If you just want to use the model rather than host it, skip all of this and use OpenRouter’s free or paid version, or try it free in OpenCode.
Open weights, open data and clear recipes make this one of the easiest big models to build on.
How I’d test it this week
If you’re evaluating self-hosting, start by testing the hosted version on your real tasks. Only if it’s good enough — and you have a privacy, cost or customisation reason — is it worth renting GPUs to host it yourself.
When you do, start with the NVFP4 checkpoint on Blackwell hardware, follow the model card’s vLLM recipe, and benchmark throughput on your own prompts before committing.
Common questions
Can I download it for free? Yes — the weights are free on Hugging Face. Can I use it commercially? The OpenMDW-1.1 licence allows commercial use, but read it yourself first. Do I need Nvidia GPUs? The official recipes target Nvidia Hopper and Blackwell GPUs; community GGUF builds run in llama.cpp-based tools like LM Studio and Ollama on other hardware. Is there a smaller version? Not of Super VL itself, but Nvidia’s Bolt 74B-A9B fits in 24GB of memory.
The bottom line on Nemotron 3.5 Super VL Hugging Face
The Nemotron 3.5 Super VL Hugging Face release is unusually open: weights, data and recipes. If you have the GPUs, it’s a strong model to host yourself; if not, use the hosted versions.
FAQ: nemotron 3.5 super vl hugging face
Is Nemotron 3.5 Super VL on Hugging Face?
Yes — BF16 and NVFP4 checkpoints from Nvidia, plus community GGUF builds.
What licence does it use?
OpenMDW-1.1, which allows commercial use.
What hardware does it need?
4× H100 or 2× GB200 for BF16; NVFP4 is more efficient.
Can I run it with vLLM?
Yes — the model card includes a vLLM recipe with tool calling.
Did Nvidia release training data?
Yes — open datasets and training recipes are on Hugging Face.
Is there a GGUF version?
Yes — community GGUF builds such as Unsloth’s are on Hugging Face, and they power the LM Studio and llama.cpp-style setups.
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 hugging face — the facts here move fast and it is updated as they do.
