Nemotron 3.5 Super VL: Specs, License & How to Run

Nemotron 3.5 Super VL is an open-weight multimodal model from NVIDIA. Official hosts we opened on 2026-10-17: the ungated card nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16, the NVFP4 sibling nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4, the OpenMDW License Agreement, the launch post From Data Center to Desktop (15 October 2026), and developer.nvidia.com Nemotron.

This identification page hangs off ChatGPT alternatives and closed LLM APIs. Those mappings are editorial. We have no sourced quality comparison against ChatGPT.

The NVIDIA blog (15 October 2026) calls it a vision-language model with 120B total / 12B active parameters and a 1M-token context, and says weights ship under the OpenMDW License 1.1 with commercial use permitted. Hugging Face pipeline_tag is image-text-to-text. The blog’s efficiency and accuracy claims are vendor claims, not a forkalt ranking.

What we verified (2026-10-17)

Field	Value	Source
HF likes / downloads	142 / 40	HF API nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16
Card created / last modified	2026-10-15T09:47:39Z / 2026-10-15T10:33:02Z	same
Gated	false	HF API
License (card)	openmdw-1.1	HF cardData.license
LICENSE file	OpenMDW License Agreement, Version 1.1	LICENSE.txt
Pipeline tag	image-text-to-text	HF API
Params (NVIDIA)	120B total / 12B active hybrid MoE	Blog + docs + HF card
Context	1,048,576 tokens	NVIDIA docs + HF card
Modalities	text + image + video in → text out	NVIDIA docs + HF card
Architecture (card)	512 routed experts per MoE layer; C-RADIO vision encoder; MTP	HF card “Model Architecture”
NVFP4 sibling likes / downloads	81 / 64	HF API …-NVFP4
NVFP4 gated	false	same
NVFP4 usedStorage	157,685,399,618 bytes (~147 GiB on disk)	HF API

The BF16 card’s usedStorage in the same API check is 239,924,488,654 bytes (~223 GiB on disk). Hugging Face safetensors.parameters on the BF16 card sum to 124,289,323,008 (~124B). We keep NVIDIA’s 120B / 12B quotes.

Hardware (vendor / datacenter-class)

runsLocally: true here means the weights are downloadable; it does not mean a consumer GPU will load them.

Copied from NVIDIA docs and the HF card (checked 2026-10-17):

Checkpoint	Deployment (NVIDIA)
NVFP4	1× DGX Spark, 1× H200, 2× H100, or 1× GB200
BF16	4× H100 or 2× GB200

Runtimes NVIDIA names: vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, Ollama, Unsloth. Hosted partners NVIDIA names: Baseten, Fireworks, Groq, Modal, Nebius, OpenRouter, Together. Those are copied names. forkalt has not opened each partner’s model SKU page.

How to run (from official docs, not yet tested)

The card and developer.nvidia.com/topics/ai/Nemotron point at self-hosting with vLLM, SGLang, or TRT-LLM, or calling a hosted provider. The model has configurable reasoning levels (off, medium, full) controlled by the chat template with enable_thinking=True/False. Recommended sampling copied from the card: temperature 1.0, top_p 0.95 for text; temperature 0.6, top_p 0.95 for vision. forkalt has not run these commands.

Example vLLM serve (BF16, 4× H100):

vllm serve nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 \
  --served-model-name nemotron-3.5-super-vl \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --gpu-memory-utilization=0.85 \
  --allowed-local-media-path=/ \
  --max-num-seqs=32 \
  --limit-mm-per-prompt '{"video": 1, "image": 1}' \
  --video-pruning-rate 0.5 \
  --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_v3 \
  --speculative_config.method mtp \
  --speculative_config.num_speculative_tokens 3

License

LICENSE.txt is titled OpenMDW License Agreement, Version 1.1, Last Updated: October 2026, copyright NVIDIA Corporation. Hugging Face cardData.license is openmdw-1.1.

Paraphrase only, not legal advice:

• Include the copyright notice and this license with copies or substantial portions.
• Prominently display “Built with NVIDIA Nemotron” attribution on products or services that use the model or derivatives.
• Commercial use including Model-as-a-Service is permitted.
• Create derivative works and distribute them under the same license.
• NVIDIA claims no ownership of outputs.

What we did not verify

Vendor benchmark tables on the card (GPQA-Diamond, HLE, MMLU-ProX, AA-Omniscience, CritPt, APEX, LiveCodeBench, and others versus named models) stay self-reported. We have no sourced bake-off against ChatGPT. We have not loaded NVFP4 or BF16 on a named cluster.

Related chat weights on this site: the compact sibling Nemotron 3.5 Bolt, DeepSeek-V4.1-Flash, MiniCPM-V 4.7, and Muse Spark Open on ChatGPT alternatives; Kolibri-1 and LFM2.5-24B-A2B on closed LLM APIs. Same Nemotron 3.5 family, different job: Nemotron 3.5 Transcribe is offline speech ASR, not this VLM. Runtimes: vLLM, DeepInfer vs vLLM.