NVIDIA Nemotron 3.5 Super VL is an open-weight vision-language model with 120B total and 12B active parameters. NVIDIA posted it on 15 October 2026. The NVFP4 card describes a Mamba-2, mixture-of-experts, and attention hybrid with a C-RADIO vision encoder and multi-token prediction. Context is up to 1M tokens. The card does not publish a single VRAM number. This calculator does not model the hybrid, so there is no preset and no will-it-fit score.
Deployment targets from the NVFP4 card
These are NVIDIA’s examples, not runs performed here. Several discrete GPUs are not one pooled memory total. DGX Spark in this catalog is one 128 GB unified-memory device. H100, H200, and GB200 are not catalog SKUs, so they are named and not linked.
| Hardware | Engine on the card | What that row says |
|---|---|---|
| 1× DGX Spark | vLLM | Text and image, 1M context, MTP with 5 draft tokens |
| 2× H100 | vLLM | Text, image, and video, 1M context, MTP with 3 draft tokens |
| 1× H200 | vLLM | Text, image, and video, 1M context, MTP with 3 draft tokens |
| 1× GB200 | vLLM | Text, image, and video, 1M context, MTP with 5 draft tokens |
| 1× GB200 or 2× H100 | SGLang | Text and image, 1M context, MTP with 5 draft steps |
| 1× GB200 | TensorRT-LLM | 1M context, MTP with 5 draft tokens |
The card also points local inference notes at DGX Spark, RTX Spark, and RTX Pro 6000. A separate LiLiCorr drafter is recommended for DGX Spark. Its bytes are not in the table below.
License: OpenMDW-1.1. Card date line: Hugging Face 15 October 2026.
Weight files
Checked 16 October 2026. GB is 10⁹ bytes. GiB is 2³⁰ bytes. Each sum is the root model-*-of-*.safetensors shards only.
| Repo | Revision | Shards | Bytes | GB | GiB | Safetensors metadata total |
|---|---|---|---|---|---|---|
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 |
579ffa0b659b |
88 | 75,010,188,240 | 75.010 | 69.859 | 65,769,840,138 |
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 |
97e009758ae2 |
63 | 248,913,061,312 | 248.913 | 231.818 | 124,453,395,976 |
The BF16 repo also contains model-summary-idxs.safetensors at 104 bytes. It is not part of the shard sum. The metadata totals are not the card’s 120B / 12B counts. The packed NVFP4 total in particular must not be read as 120 billion parameters.
File size is not resident memory. See file size is not runtime VRAM. The smaller packed sibling aimed at a 24 GB label is Nemotron 3.5 Bolt.
What this page does not claim
The blog’s throughput wording is NVIDIA’s claim. It is not repeated here as a measurement. No calculator scenario covers Super VL. Qwen3.8-27B is a catalog model the estimator does cover. Formula limits are on the methodology page.
FAQ
How much VRAM does Nemotron 3.5 Super VL need?
The NVFP4 model card does not publish one VRAM number. It lists deployment targets: 1x DGX Spark, 1x H200, 2x H100, or 1x GB200, with the modality and context depending on the recipe. The NVFP4 safetensors shards sum to 75,010,188,240 bytes (75.010 GB, 69.859 GiB). The BF16 shards sum to 248,913,061,312 bytes (248.913 GB, 231.818 GiB). Those are file sizes, not a measured resident set.
Is the NVFP4 parameter metadata the model size?
No. The card says 120B total and 12B active. The NVFP4 safetensors metadata total is 65,769,840,138, a packed count. The BF16 metadata total is 124,453,395,976. Use the card's 120B / 12B figures for the model scale.
What context length does Super VL support?
The card says up to 1M tokens. The prompt, visual inputs, history, tool results, and the reply share that window. Individual recipes may use less context so the GPU has room for concurrency.
Are the Super VL weights open?
Yes. NVIDIA's 15 Oct 2026 blog says Super VL is available on Hugging Face, OpenRouter, the ModelScope community, and as NVIDIA NIM. The checked weights use the OpenMDW-1.1 license. The card's Hugging Face date is 15 Oct 2026.