The official Nemotron 3.5 Super VL Hugging Face repo is nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. That's where NVIDIA publishes the open weights, under the OpenMDW License Agreement v1.1. Grab it there if you want to serve it yourself, fine-tune it or build on it. If you just want to use it, skip the download and run it hosted.
📺 Watch: NVIDIA Nemotron 3.5 Super VL Is INSANE! (FREE AI Model)
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
This page is for builders. I'll explain what's actually in the release, what the repo name means, what the licence lets you do in plain English, and which tools NVIDIA names for serving and fine-tuning. Hardware warning up front: this is a 120B model and it needs big-memory hardware to run.
If you'd rather plug open models into a ready-made agent setup than build from raw weights, my AI Profit Boardroom ($69/mo) includes the Agent OS zip, a video tutorial and a 30-day roadmap.
Where to Find Nemotron 3.5 Super VL Hugging Face Weights
The repo is nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4. Always grab it from the nvidia account so you know you're getting the real thing, not a re-upload.
The name tells you a lot once you break it down:
- Nemotron-3.5-Super-VL: the model, VL meaning vision-language
- 120B: 120 billion total parameters
- A12B: 12 billion active per token, because it's mixture-of-experts
- NVFP4: NVIDIA's 4-bit quantized format
The weights are also on ModelScope if that suits your setup better. For the full rundown on what the model does, see my Nemotron 3.5 Super VL overview.
What's in the Release
NVIDIA went further than just dropping weights. Per NVIDIA, the release includes:
- Base checkpoint: the pre-trained model before instruction tuning. Use this if you want to do your own post-training from scratch.
- Post-trained checkpoint: the version tuned for reasoning, instruction following and tool use. This is the one most people want.
- NVFP4-quantized checkpoint: the 4-bit version, built with quantization-aware distillation to keep quality while cutting size.
- Training recipes, data and RL environments: how they built it, so you can reproduce or extend it.
- Open datasets on Hugging Face: including 40 vision-language datasets.
- A usage cookbook on GitHub: examples for getting it running.
That's a lot more open than most model drops. You're not just getting a finished model. You're getting the ingredients.
What NVFP4 means for your hardware
NVFP4 runs at 4-bit on NVIDIA Blackwell GPUs. Per NVIDIA, the same checkpoint also works on non-Blackwell hardware using 4-bit weights with 16-bit activations. So you don't need Blackwell to use it, but Blackwell is where it's designed to shine.
Either way, you need a lot of memory. Local runs need roughly 70-90GB, so think a 128GB DGX Spark or a big-memory workstation. Not a laptop.
📺 Watch: OpenClaw + Nvidia Nemotron 3 Super + Ollama is INSANE!
The Licence in Plain English
The model ships under the OpenMDW License Agreement v1.1. It's an open-weights licence. In plain terms, NVIDIA is letting you download, run and build on the weights.
But here's my honest advice: read the licence yourself before you ship anything commercial. I'm not a lawyer and I'm not going to pretend I can tell you every condition in it. Open weights doesn't automatically mean "do anything you like with zero conditions". The licence text sits with the model on Hugging Face. Ten minutes reading it beats a headache later.
The Open Datasets
This is the part builders should get excited about. NVIDIA put open datasets on Hugging Face alongside the model, including 40 vision-language datasets.
Why that matters:
- You can see what kind of data shaped the model's vision skills
- You can reuse them to fine-tune your own models
- You can mix them with your own data for a domain-specific version
Pair that with the training recipes and RL environments, and you've got a real starting kit for building your own vision agent.
📺 Watch: New FREE Nvidia Nemotron 3 Super + Ollama
Serving It: vLLM, SGLang and NIM
Once you have the weights, you need something to serve them. NVIDIA names these options:
- vLLM: a popular open-source inference server, listed as an inference partner
- SGLang: another open-source serving framework, also an inference partner
- NVIDIA NIM: NVIDIA's packaged inference microservices
- NVIDIA Dynamo: recipes for scaling across many GPUs
I'm not going to make up launch commands here. The settings depend on your GPU, your memory and which checkpoint you pick. Use the usage cookbook on GitHub and each tool's own docs for the exact setup.
For scale, NVIDIA reports that with Dynamo on 16 B200 GPUs at 50 tokens per second per user, it served up to 8x as many users with 85% lower time to first token. That's NVIDIA's own number on serious data-centre hardware, with no independent replication yet.
If you'd rather skip serving entirely, Ollama, LM Studio, Unsloth and Exo are listed as inference partners too, and clouds like Amazon SageMaker JumpStart, Microsoft Foundry, Baseten, Fireworks AI and Together AI host it.
Not sure whether to self-host or go hosted for your business? Book a free AI strategy session and we'll work out what makes sense.
Fine-Tuning: The Tools NVIDIA Names
NVIDIA points to these for customising the model:
- NeMo AutoModel: for supervised fine-tuning (SFT) and LoRA
- NeMo RL: for reinforcement learning on top of the model
- Unsloth: a popular fine-tuning toolkit
LoRA is the realistic starting point for most people. You train a small set of extra weights instead of the whole model, which needs far less compute. Full fine-tuning a 120B model is a data-centre job.
Good fine-tuning targets for a vision model like this:
- Reading your company's specific document formats
- Understanding charts and dashboards in your industry
- Pulling fields out of invoices or forms you process every day
A Simple Builder Checklist
- Decide your goal: use, self-host, fine-tune or build a product.
- Check your hardware honestly. Roughly 70-90GB of memory for local runs, more for long context and images.
- Pick the checkpoint: post-trained for most jobs, base for your own post-training, NVFP4 for a smaller footprint.
- Read the OpenMDW-1.1 licence before anything commercial.
- Pick a serving tool (vLLM, SGLang or NIM) and follow its docs plus NVIDIA's cookbook.
- Test it on your own real files before you trust it. Vendor numbers are a starting point, not proof.
Do You Actually Need the Raw Weights?
Probably not. Be honest about your goal.
If you want to use the model, the hosted route is way easier. It's on OpenRouter as nvidia/nemotron-3.5-super-vl with a :free variant. I've used it free inside OpenCode in my Agentic Operating System, picking "Nemotron 3.5 Super VL" in the drop-down. I asked it to build a beautiful to-do list app, make it entertaining, engaging, and the build came out working.
My take: not a frontier model, not going to replace Opus 5.5. But free, open, fast, and good at reading screenshots, PDFs, charts and receipts.
Grab the Hugging Face weights if you want to fine-tune, self-host for privacy, or build a product on top. If you want to compare how NVIDIA itself positions the model, read my NVIDIA Nemotron 3.5 Super VL breakdown. And if you want a click-and-go local option instead, my LM Studio guide for Nemotron 3.5 Super VL covers the 70GB Q4_K_M build.
For how I test models on real jobs rather than vendor charts, see Goldie Bench. And to wire open models into agents, my Agent OS guide shows how I run Hermes, OpenClaw, OpenCode and Claude side by side.
Inside the AI Profit Boardroom, 3,400+ members get 4 weekly coaching calls and daily tutorials on setups like this. Or grab a free AI strategy session to talk through your build.
FAQ
What is the official Hugging Face repo for Nemotron 3.5 Super VL?
nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4, published under the nvidia account.
Is Nemotron 3.5 Super VL open source?
It's open weights under the OpenMDW License Agreement v1.1, and NVIDIA also released training recipes, data and open datasets. Read the licence before commercial use.
Can I fine-tune it?
Yes. NVIDIA names NeMo AutoModel for SFT and LoRA, NeMo RL for reinforcement learning, and Unsloth. LoRA is the practical route for most people.
Do I need a Blackwell GPU?
No. NVFP4 runs at 4-bit on Blackwell, and per NVIDIA the same checkpoint runs on non-Blackwell hardware with 4-bit weights and 16-bit activations. You still need a lot of memory.
