2026-10-16
Nemotron 3.5 Super VL: 120B multimodal on one GPU
NVIDIA's new open model reads text, images and video with 12 billion active parameters. Here is the hardware it actually needs, from one H200 to an RTX PRO 6000.
Nemotron 3.5 Super VL is the vision-language model NVIDIA published on 15 October 2026 for coding agents, document research, and image and video understanding. It joins the SophyAI catalog with three reference profiles.
Architecture
It is a hybrid Mamba-2 and Transformer Mixture of Experts. It has 120 billion parameters and activates 12 billion per token. Each MoE layer contains 512 experts. A C-RADIO vision encoder and a connector bring images, pages and frames into the language backbone. Output is text only.
| Feature | Value |
|---|---|
| Total / active parameters | 120B / 12B |
| Input | text, images, video |
| Output | text |
| Native context | up to 1,048,576 tokens |
| Speculative decoding | native Multi-Token Prediction |
| License | OpenMDW-1.1 |
Reasoning is configurable: off, medium or on. For text, NVIDIA recommends temperature 1.0 and top_p 0.95; for vision, temperature 0.6 and the same top_p. Published videos run up to 2 minutes. The vLLM recipe samples up to 256 frames at 2 FPS for 720p, and 1 FPS with 128 frames for 1080p.
Two checkpoints, two footprints
The NVFP4 checkpoint occupies 75 GB on Hugging Face. Vision, connector, Mamba convolutions and embeddings stay in BF16; the router stays in FP32. NVIDIA specifies W4A4 inference on Blackwell and W4A16 on Hopper.
The BF16 checkpoint occupies 249 GB and is the reference for customization and full-precision evaluation. Its published deployment uses 4× H100 or 2× GB200.
Hardware for NVFP4 inference
NVIDIA publishes these vLLM configurations, all with up to 1M context:
| Hardware | Modalities | MTP |
|---|---|---|
| 1× H200 | text, images, video | 3 tokens |
| 2× H100 | text, images, video | 3 tokens |
| 1× GB200 | text, images, video | 5 tokens |
| 1× DGX Spark | text and images | 5 tokens |
SGLang covers text and images on 1× GB200 or 2× H100. TensorRT-LLM is documented on 1× GB200 for 1M tokens. On DGX Spark, its recipe drops to 131,072 tokens.
For workstations, NVIDIA points to llama.cpp and Unsloth on DGX Spark, RTX Spark and RTX PRO 6000. The Unsloth NVFP4 GGUF weighs 77.7 GB across three shards, plus a 1.7 GB vision projector. It fits in 96 GB of VRAM, but leaves little headroom: a 1M context should not be promised on this GPU until measured. NVIDIA does not publish tok/s for any of these configurations.
SophyAI profiles
Three recipes are visible and not yet launchable:
- vLLM NVFP4 on 1× H200, 1M, text, images and video.
- vLLM NVFP4 on 2× H100, with the same coverage.
- llama.cpp NVFP4 on 1× RTX PRO 6000, with a cautious 131,072-token cap.
They are reference: we do not publish SophyAI prices or benchmarks. The repository is not gated.
For a deployment today, I would choose one H200 when a single data-center GPU is required. The H100 pair is the official Hopper alternative. RTX PRO 6000 is the workstation path and its context must be measured first. BF16 remains reserved for customization and evaluation.