Nvidia introduces Nemotron 3.5 Super VL for enterprise AI agents

Nvidia has released Nemotron 3.5 Super VL, an open 120-billion-parameter model designed to help enterprise AI agents work with documents, charts, tables, images and webpages. Carl Franzen reports for VentureBeat that the model is aimed at tasks where an agent must combine visual information, written instructions, company policies and external tools.

The launch adds visual understanding to Nvidia’s larger Nemotron Super model family. Companies can further train Super VL on their own data, procedures and tools, then run it in cloud environments or on infrastructure that meets their security and data-residency needs.

Nvidia says partners are adapting the model for financial document reviews, insurance claims, cybersecurity, scientific work and coding. The company’s benchmark chart shows specialized, post-trained versions improving by 9% to 44% over the base model on five tasks. In some of those narrow evaluations, Nvidia says the customized models score higher than named Claude and GPT models.

Those results should not be read as a general comparison between base models. Nvidia has not published the underlying task definitions, sample sizes or test protocols. Organizations would need to validate the model against their own documents, workflows and accuracy requirements.

Nvidia also highlights generation speed. It says Super VL reaches 625 output tokens per second on four Blackwell B200 GPUs under a test with 100,000 input tokens. That measures model output, not the full time an agent needs to read files, use external tools and check its work.

Two smaller Bolt models target local deployment. Bolt 74B-A9B activates 9 billion parameters at once and is intended to run in 24GB of memory for coding, research and other multistep work. Bolt E4B, due October 23, targets smaller device-based assistants.

Super VL’s weights are available through Hugging Face, while Nvidia offers a free rate-limited development API. For production deployments, Nvidia lists its AI Enterprise software license at $4,500 per GPU annually, excluding hardware, hosting and operational costs.