One open model reads the receipt and the table

NVIDIA has released Nemotron 3.5 Super VL, an open multimodal model it describes as built for specialized agent workflows. The company announced it on 15 October 2026, and developer.nvidia.com presents it as a single model that handles reasoning, instruction following and tool use alongside text, image, document and video understanding.

The architecture is a mixture of experts with 120B parameters in total, of which 12B are active per token. That split is the usual way to keep a large model's capacity while paying for a smaller amount of computation on each step. NVIDIA reports up to 5x higher throughput on agentic tasks than other open models in the same size class, measured on a single GPU.

Ramp turns receipt images into JSON

The concrete example on the page is Ramp, a financial operations platform provider, which is converting receipt images into structured JSON holding merchant information, dates and payment data. Ramp expects up to 50% higher throughput and 10x lower latency from the post-trained model. Scale AI post-trained it for automated vendor credit processing at Skydance, and the page states inference cost for that workload will fall by 40x; the same passage gives 5x lower latency, while the page's summary block at the top says 10x. AI21 and Alice used it for threat intelligence and vulnerability analysis, and the body text gives a 41% accuracy improvement, against 40% in the summary block.

Those figures are the organizations' own results for their post-trained versions, not independent measurements of the base model. What the release changes for document work is the packaging: one model reads a page layout, a table and a chart and reasons over them, so a pipeline does not need a separate text model and vision-language model with task context passed between them. For incoming invoices and contracts, that is the shape of the job — read the page, find the fields, return them in a fixed structure.

NVIDIA says the model's visual training covers charts, tables, optical character recognition, multipage documents and video. On CharXiv, a chart-reasoning benchmark, it scored 80.8%, against 63% for Nemotron 3 Nano Omni, and 84% with Python code interpretation. NVIDIA also states that it is one of the top open models on the OCRBench v2 leaderboard.

Open weights and datasets for post-training

Developers can adapt the weights with NVIDIA NeMo using supervised fine-tuning, low-rank adaptation and reinforcement learning, and deploy them locally, in a data center or in the cloud. Alongside the model, NVIDIA is releasing open datasets on Hugging Face covering agentic tool use, finance reasoning and tool calling, instruction following, long-context work and 40 vision-language datasets.

developer.nvidia.com