Nemotron 3.5 Super VL: NVIDIA's Open 120B Model for Agents

NVIDIA has released Nemotron 3.5 Super VL, an open multimodal model built to sit at the center of AI agents that need to read more than plain text. Announced on October 15, the model takes in text, images, documents and video, then reasons over them, follows instructions and calls tools, all inside a single network. For teams building document-processing or research agents, that means one open model can replace a stack of separate vision and language systems.

• Size: 120B total parameters in a mixture-of-experts design, with about 12B active per token.
• Inputs: text, images, documents and video, plus tool use and reasoning in the same model.
• Released: base, post-trained and NVFP4 checkpoints, training recipes, reinforcement-learning environments and 40 vision-language datasets.
• Where to run it: Hugging Face, OpenRouter, NVIDIA NIM, DGX Spark and major clouds, according to NVIDIA.

What Is Nemotron 3.5 Super VL?

Nemotron 3.5 Super VL is the vision-language member of NVIDIA's Nemotron 3.5 family. Like the Nemotron 3.5 Bolt models we covered yesterday, it uses a hybrid mixture-of-experts architecture, so only a fraction of its 120 billion parameters fire for any single token. That keeps inference costs closer to a 12B model while retaining the knowledge capacity of a much larger one.

Where Bolt targets local hardware, Super VL is aimed at production agent workloads: reading charts, parsing receipts and invoices, understanding screenshots, and turning messy visual inputs into structured output that downstream tools can use.

How Does Super VL Perform?

NVIDIA's own evaluations, which have not yet been independently reproduced, point to strong gains over earlier Nemotron releases:

• Chart reasoning (CharXiv): 80.8%, rising to 84% when the model can run Python, compared with 63% for Nemotron 3 Nano Omni.
• Agentic tool use: 63.44% on Hermes Bench and 87.05% on LangChain Evals, both ahead of Nemotron 3 Ultra in NVIDIA's testing.
• Throughput: NVIDIA says the model delivers up to 5x higher agentic throughput on a single GPU than open models of similar size, and more than 600 output tokens per second on a Baseten endpoint.

VentureBeat, which reviewed the release, notes that headline wins against closed frontier models come from customized versions on specific tasks rather than across-the-board results. That is a sensible way to read the launch: Super VL is a strong open foundation designed to be specialized, not a general chatbot replacement.

Why Open Training Data Matters for Agents

The most useful part of this release may be everything shipped alongside the weights. NVIDIA published the training recipes, reinforcement-learning environments and dozens of datasets covering agentic tool use, finance, long context and vision-language tasks. It also released an Indic dataset of roughly 2.8 trillion tokens across 22 Indian languages, built with IIT Madras and AI4Bharat.

Early partner results show what specialization can unlock. NVIDIA reports that Ramp saw up to 50% higher throughput turning receipts into structured JSON, and that Scale AI cut inference cost for a document-processing workflow by 40x. For enterprise teams, open data plus open recipes means a fine-tuned agent can be audited and rebuilt rather than treated as a black box.

Where Can Developers Try It?

The model is available today on Hugging Face, ModelScope and OpenRouter, as an NVIDIA NIM microservice, and on Amazon SageMaker JumpStart, Google Cloud, Microsoft Foundry and Oracle Cloud. Inference providers including Baseten, CoreWeave, Together AI and Fireworks AI host it as well, and it runs in vLLM and SGLang for self-hosting. NVIDIA's NeMo tools handle fine-tuning, LoRA and reinforcement learning for teams that want to adapt it.

Open multimodal models have been closing the gap with proprietary systems all year, and Super VL pushes that trend squarely into agent territory. For more on where model development is heading, browse our AI coverage.

Sources: NVIDIA Technical Blog, NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning — October 15, 2026; VentureBeat, Nvidia's new open Nemotron 3.5 Super VL — October 15, 2026.