NVIDIA Nemotron 3.5 Super VL: What's New, Benchmarks, and How to Run It

NVIDIA just released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model built for always-on AI agents. It combines reasoning, instruction following, and tool use with text, image, document, and video understanding in a single checkpoint. This guide covers what the model is, how it performs, how its licensing works, and how to run it — so you can decide whether it belongs in your agent stack.

What is Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL is NVIDIA's open, customizable vision-language model for specialized enterprise agents. Its 120B-parameter mixture-of-experts architecture activates 12B parameters per token for agentic workflows. That design balances capacity with inference cost: you get a large model's knowledge without paying to run all 120B parameters on every token.

The headline feature is consolidation. The model combines reasoning, instruction following, tool use, and text, image, document, and video understanding in a single model, eliminating the need for separate vision-language models. For developers, that means one open checkpoint replaces a separate text model along with a separate vision model.

It landed on October 15, 2026 as part of NVIDIA's broader Nemotron family. Super VL and Bolt 74B-A9B are available now, with Bolt E4B coming Oct. 23. The three open models together cover the full range of agent deployments, from large enterprise servers down to edge devices and local hardware.

Architecture: hybrid Mamba-Transformer MoE

Under the hood, Super VL uses an efficiency-first design. Nemotron 3.5 Super VL builds on the Nemotron 3 Super architecture with a 120B-parameter Mixture-of-Experts model that activates 12B parameters per token. Its hybrid Mamba-Transformer architecture supports long-context processing, while multi-token prediction (MTP) helps accelerate generation.

The model wasn't trained from scratch. NVIDIA distilled Super VL from Nemotron 3 Ultra, its largest reasoning model. Training was substantial: NVIDIA reports the model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments.

Several inference optimizations stack on top:

• LiLiCorr speculative decoding, which NVIDIA says delivers up to 2.3x faster token generation than decoding without speculation on DGX Spark.
• NVFP4 quantization with quantization-aware distillation, so the 4-bit checkpoint keeps accuracy close to the original BF16 model.
• Context reuse across agent turns via prefix caching, preserving both attention caches and Mamba recurrent states.

NVIDIA also built the model with outside help. NVIDIA developed the model with the Nemotron Coalition, a group that includes Mistral AI, Thinking Machines Lab, Perplexity, and Black Forest Labs. High-quality datasets from Perplexity helped improve and streamline the model's mid-training. Environments from Prime Intellect enabled building better teacher models.

Benchmarks: speed and accuracy

NVIDIA's core efficiency claim is throughput. On a single GPU, Nemotron 3.5 Super VL delivers up to 5x higher throughput for agentic tasks compared to other open models in its size category. That speed compounds into faster task completion for long-running agents, which matters when a single job spans hours or days.

On agent-behavior benchmarks, it beats NVIDIA's own larger model:

Those figures come from NVIDIA's technical blog, which also notes that Nemotron 3.5 Super VL is also one of the top open models on the OCRBench v2 leaderboard. On chart reasoning specifically, NVIDIA reports the model scored 80.8% on the CharXiv benchmark, rising to 84% with Python code interpretation. On coding, NVIDIA says the model outperforms Nemotron 3 Super on SWE-bench and Terminal-Bench across multiple evaluation setups.

Independent coverage frames the accuracy claim more pointedly: VentureBeat reports that, with post-training, the model is positioned to beat frontier models like Claude Opus 5 and GPT-6.1 Sol on specialized task benchmarks. That caveat matters — the frontier-matching claims hinge on fine-tuning for a specific workflow, not out-of-the-box general use.

As always with vendor benchmarks, treat these numbers as NVIDIA's own measurements until third parties reproduce them. One independent signal exists: NVIDIA cites an EXO Labs evaluation that placed Super VL on the Pareto frontier for accuracy and task completion time on DGX Spark, per its technical blog.

What it actually does well

NVIDIA targets the model at work agents encounter in practice. For vision, NVIDIA focused on the kind of material agents encounter in daily work, including OCR, tables, charts, multi-page documents, and videos. The flagship example is cross-referencing: the model can interpret a chart, connect it with a table elsewhere in a document, and summarize both.

The documented agent workflows span coding, web research, document intelligence, cybersecurity triage, and video understanding. Early adopters report concrete gains after post-training:

• Ramp is converting receipt images into structured JSON and expects up to 50% higher throughput and 10x lower latency.
• Scale AI post-trained the model for vendor credit processing, projecting 40x lower inference cost and 5x lower latency on that workload.
• AI21 and Alice post-trained it for threat intelligence and vulnerability analysis, improving accuracy by 41%.

These are early, workload-specific results from NVIDIA's launch materials, not general benchmarks — read them as what the model can reach with tuning, not guaranteed defaults.

Nemotron 3.5 Bolt: the edge models

Super VL doesn't ship alone. Two smaller Bolt models bring agents to local hardware:

• Nemotron 3.5 Bolt 74B-A9B 24GB targets local coding, research and other multistep tasks. It uses a native 2-bit quantization to fit in 24GB and run alongside other apps — NVIDIA calls it the #1 open model in its class on PinchBench and #2 for BrowseComp.
• Bolt E4B 4GB is built for conversational and physical agents on GeForce RTX, Jetson, and RTX/DGX Spark, for fast local voice assistants and embedded interfaces.

Together, the lineup covers data center (Super VL), on-device multistep work (Bolt 74B), and lightweight voice (Bolt E4B).

Licensing: is it really open?

Yes, with the usual caveats. The model ships with open weights, open datasets, and training recipes. According to reporting around the launch, Nemotron 3.5 Super VL is distributed under NVIDIA's OpenMDW license, which is described as permitting commercial use, fine-tuning, and redistribution. That positions it closer to permissive open-source terms than the restrictive research-only licenses some labs still attach to their largest models.

The practical catch is hardware. Dozens of organizations are already fine-tuning Super VL, but getting the most out of it means buying NVIDIA hardware. The optimized NVFP4 path and most of the serving tricks are tuned for NVIDIA's Blackwell and DGX stack. Always confirm the exact terms on the model card before shipping commercially.

How to run Nemotron 3.5 Super VL

The release includes base, post-trained, and NVFP4-quantized checkpoints. You have three broad paths:

1. Self-host. Grab the BF16 or NVFP4 safetensors and serve with vLLM, SGLang, or TensorRT-LLM. The NVFP4 checkpoint is designed to run on a single H200 or DGX Spark.
2. Hosted API. Super VL is available via NVIDIA NIM, OpenRouter, Together AI, Baseten, and fal, with a free tier on build.nvidia.com.
3. Local / edge. Use the Bolt 74B-A9B model through Ollama or LM Studio for workloads you want to keep off the cloud.

For customization, NVIDIA points developers to NeMo for supervised fine-tuning, LoRA, and reinforcement learning, plus open datasets on Hugging Face. The full architecture breakdown lives in NVIDIA's technical blog and launch announcement.

Should you use it?

Nemotron 3.5 Super VL is a strong fit if you run agents that read messy visual documents, you want to keep data on your own infrastructure, and you have a specific workflow worth post-training for. It's less compelling if you need a general-purpose chat model or you aren't on NVIDIA hardware — the efficiency story leans heavily on that stack.

For a sense of how it sits against other recent open and frontier launches, see our breakdowns of Gemini 4 Argon and GLM-5.3-Flash.

Put an open multimodal model to work

Nemotron 3.5 Super VL is built to read documents, call tools, and run multistep tasks — exactly the work an AI workforce should handle on its own. Eigent is an open-source Cowork desktop app that orchestrates multi-agent workflows locally, so you can point agents at your own documents and tools instead of stitching models together by hand. See how it handles document-heavy review workflows, or download Eigent and build your own agent team.