# NVIDIA’s Nemotron 3.5 Super VL: A Complete Guide to the Open Multimodal Agent Model
## What Is Nemotron 3.5 Super VL?
NVIDIA has unveiled Nemotron 3.5 Super VL, a large-scale open-source model designed for enterprise agents that need to work continuously with multiple types of input. At 120 billion parameters total, the model uses a mixture-of-experts design, meaning only about 12 billion parameters are activated for each token it processes. This makes it efficient enough to run on a single high-end GPU while still delivering powerful multimodal capabilities.
The model handles text, images, document pages, and video clips simultaneously. It can reason through complex tasks, write code, make tool calls, and generate structured outputs — all within one unified system. For teams that previously needed a separate model for language tasks and another for vision tasks, this single checkpoint replaces both.
The weights are released under the OpenMDW-1.1 license, and NVIDIA has also published open datasets, training recipes, and reinforcement learning environments through its NeMo Gym framework to help developers fine-tune and extend the model.
## Architecture: How It Works Under the Hood
The backbone of Nemotron 3.5 Super VL combines three types of layers in sequence: Mamba-2 layers, mixture-of-experts layers, and attention layers. Each experts layer routes tokens across 512 specialized experts, and a vision component called C-RADIO encodes images, document pages, and video frames before feeding them into the language backbone.
NVIDIA focused on four techniques to make the model fast:
– **Multi-token prediction:** A built-in drafter suggests several tokens at once and verifies them in a single forward pass, reducing the number of computation steps needed per generated token.
– **Speculative decoding (LiLiCorr):** A lightweight draft model running on NVIDIA’s DGX Spark hardware accelerates token generation by up to 2.3 times.
– **Smart quantization:** The NVFP4 checkpoint uses 4-bit weights and activations on Blackwell GPUs, while using 4-bit weights with 16-bit activations on other hardware. This keeps quality high while drastically reducing memory use.
– **Context reuse:** Prefix caching preserves both attention caches and Mamba states between agent interactions, so repeated work is avoided during ongoing conversations.
The model was distilled from the earlier Nemotron 3 Ultra and trained on roughly 26 trillion tokens. Its post-training phase used about 3.8 million samples combined with reinforcement learning across dozens of simulated environments. An additional round of continued pretraining introduced 10 billion tokens in Indian languages.
## Benchmarks: Where It Excels and Where It Falls Short
NVIDIA reports the following performance scores for the NVFP4 version of the model:
| Area | Benchmark | Score |
|——|———–|——-|
| Reasoning | GPQA-Diamond | 89.90 |
| Coding | SWE-bench Verified | 71.07 |
| Coding | Terminal-Bench 2.1 | 60.80 |
| Web Agent | BrowseComp | 77.5 |
| Instruction Following | IFBench | 71.50 |
| Charts | CharXiv Reasoning | 79.88 |
| Documents | MMLongBench-Doc | 62.82 |
| Computer Use | OSWorld-Verified | 52.3 |
| Long-Horizon Agent | Tau3-Bench Banking | 17.78 |
In head-to-head comparisons against the older Nemotron 3 Ultra, Super VL outperformed its predecessor on agent harness evaluations, scoring 63.44% on Hermes Bench compared to 56.33% for Ultra, and 87.05% on LangChain-style evaluations versus 84.3%. Notably, Super VL achieves these results while having roughly 4.6 times fewer total parameters.
NVIDIA also reports over 600 output tokens per second on agentic task benchmarks through a Baseten hosting endpoint, claiming up to five times the single-GPU throughput of comparable open-source vision-language models.
## Deployment Options
Developers can access Nemotron 3.5 Super VL through several platforms including Hugging Face, OpenRouter, ModelScope, and as an NVIDIA NIM container. Supported inference runtimes include vLLM, SGLang, and TensorRT-LLM on both Blackwell and Hopper GPU architectures, with local support through llama.cpp and Unsloth.
For large-scale production, NVIDIA’s Dynamo framework offers KV-aware routing that, on a 16-GPU B200 setup, reportedly serves eight times more users at 50 tokens per second per user while cutting first-token latency by 85%.
Customization is supported through NeMo AutoModel for supervised fine-tuning and LoRA adapters, as well as NeMo RL for reinforcement learning. Early production deployments include vendor credit processing that achieved 40 times lower inference cost, receipt-to-JSON extraction with significantly higher throughput, and threat intelligence analysis that improved accuracy by 40%.
## The Nemotron 3.5 Bolt Family
Alongside the main Super VL model, NVIDIA released two smaller edge-optimized variants:
**Nemotron 3.5 Bolt 74B-A9B** is designed for local agent workloads. It was distilled from Super VL and can run natively in 2-bit quantization within 24 gigabytes of memory, making it compatible with consumer RTX GPUs and DGX Spark systems. It currently holds the top ranking among open models in its size class on the PinchBench leaderboard and ranks second on BrowseComp.
**Nemotron 3.5 Bolt E4B (4GB)** is a compact model targeting voice assistants and embedded agents on GeForce RTX, Jetson, and similar edge devices.
Both Bolt models are available on Hugging Face and support popular frameworks including llama.cpp, Ollama, LM Studio, Unsloth, and vLLM.
## Frequently Asked Questions
**Q: What does the “VL” in Nemotron 3.5 Super VL stand for?**
A: It stands for “Vision-Language,” reflecting the model’s ability to process both visual inputs (images, documents, video) and text.
**Q: Can I run this model at home?**
A: The full 120B model requires professional hardware such as a DGX Spark, H200, or GB200. However, the smaller Bolt 74B-A9B model can run on consumer RTX GPUs with 24GB or more of VRAM.
**Q: What license is the model released under?**
A: The weights are available under the OpenMDW-1.1 license, which is designed for open research and commercial use.
**Q: How long of a video can the model process?**
A: Video input is limited to MP4 clips up to 2 minutes in length.
**Q: Does the model support chain-of-thought reasoning?**
A: Yes. Reasoning can be toggled on or off through the chat template, giving developers control over whether the model shows its internal reasoning steps.
**Q: What sets this apart from models like Qwen3.5-122B?**
A: Nemotron 3.5 Super VL is specifically designed as an always-on enterprise agent with integrated vision capabilities, native document and video understanding, and aggressive quantization for deployment efficiency.
**Q: Are there community resources for learning how to use it?**
A: Yes. NVIDIA has published a technical report, a detailed model card, and a usage cookbook to help developers get started.
## Conclusion
Nemotron 3.5 Super VL represents a significant step forward in making powerful multimodal agent models accessible to the broader developer community. By combining strong reasoning scores, native support for images and video, and efficient deployment through quantization and speculative decoding, NVIDIA has created a model that balances capability with practicality. The addition of the Bolt family extends this accessibility to edge and local devices, making it possible to run capable agents on consumer hardware as well. As open-source multimodal models continue to evolve, Super VL sets a new benchmark for what a single model can handle across text, vision, and agentic tasks.
Thank you for reading
