NVIDIA has announced Nemotron 3.5 Super VL, an open multimodal model designed for long-running AI agents.
- 120B total parameters
- 12B active parameters
- Mixture-of-Experts architecture
- context window up to 1M tokens
- support for text, images, documents, and video
- reasoning and tool use
- C-RADIO vision encoder
- BF16 and NVFP4 checkpoints
- support for vLLM, SGLang, and TensorRT-LLM
120B total parameters, 12B active. The model does not use all 120B parameters for every token. About 12B parameters are active during inference, reducing compute requirements compared with a dense 120B model.
Technical specifications
| Specification | Nemotron 3.5 Super VL |
| Total parameters | 120B |
| Active parameters | 12B |
| Architecture | Mamba-2 + Transformer + MoE |
| Context window | Up to 1M tokens |
| Vision encoder | C-RADIO |
| Input | Text, images, documents, video |
| Output | Text |
| Tool calling | Yes |
| Reasoning | Yes |
| Quantization | NVFP4 |
| Full precision | BF16 |
Nemotron 3.5 Super VL vs Nemotron 3 Super
The main change is multimodality.
| Feature | Nemotron 3 Super | Nemotron 3.5 Super VL |
| 120B total parameters | Yes | Yes |
| 12B active parameters | Yes | Yes |
| 1M context | Yes | Yes |
| MoE | Yes | Yes |
| Mamba + Transformer | Yes | Yes |
| Tool use | Yes | Yes |
| Reasoning | Yes | Yes |
| Image understanding | No native VL stack | Yes |
| Document understanding | External pipeline | Yes |
| Video understanding | No | Yes |
| Vision encoder | No | C-RADIO |
Nemotron 3.5 Super VL is not mainly a larger model. It extends the Nemotron 3 Super architecture with native vision capabilities.
What the model can process
Nemotron 3.5 Super VL is designed for agent workloads involving multiple data types:
- OCR
- tables and charts
- screenshots
- document layouts
- images
- short video clips
- tool calls
- long conversation history
- coding and terminal-agent tasks
For video, the model works with sampled frames rather than continuous unrestricted video input.
Architecture
A simplified pipeline looks like this:
Image / Video → C-RADIO → Vision Connector → Mamba-2 + Transformer + MoE → Text
For text-only input, the vision encoder is skipped.
The architecture combines:
- Mamba-2 for efficient long-sequence processing
- Transformer attention for precise context handling
- MoE for large model capacity with limited active compute
1M-token context
The 1M-token limit is shared across the entire request.
It includes:
- system instructions
- conversation history
- documents
- visual tokens
- tool outputs
- generated output
Using the full context window also increases memory usage and latency, so production deployments may use smaller limits.
BF16 and NVFP4
NVIDIA provides both:
- BF16 for full-precision deployment.
- NVFP4 for lower-memory inference.
Not every component is quantized to 4-bit precision. Some parts, including the vision stack and routing components, remain at higher precision.
Hardware requirements
The 12B active-parameter figure should not be confused with a normal dense 12B model. The full 120B parameter set still has to be stored or distributed across GPUs.
Reference configurations include NVIDIA H100, H200, GB200, Blackwell GPUs, and DGX Spark for reduced-context deployments.
What changes for developers
A multimodal agent pipeline might previously look like this:
OCR → vision model → LLM → tools
With Nemotron 3.5 Super VL, some of that can be consolidated into:
Nemotron 3.5 Super VL → tools
The main upgrade over Nemotron 3 Super is therefore not model size.
It is the combination of reasoning, tool use, image understanding, document processing, and video understanding in one model.
