Nvidia's open Nemotron 3.5 Super VL reads receipts, charts and PDFs for agents. Ramp is using it for receipt data

Document intake is the most common automation I build, and Nvidia just released an open model aimed straight at it.

What happened

Nvidia released Nemotron 3.5 Super VL, an open multimodal model with 120 billion parameters, of which about 12 billion are active per token. Nvidia built it for specialized agents that run continuously and handle text, images, documents and video.

The vision work targets what agents see in daily work: OCR, tables, charts, multi-page documents and web pages. Nvidia says the model can read a chart, link it to a table elsewhere in the same document and summarize both without a separate image model. Nvidia reports 80.8% on the CharXiv chart benchmark. The weights ship with open datasets and training recipes, and Nvidia pitches it to companies that want to fine-tune on their own data and run it on their own infrastructure.

Nvidia says dozens of organizations are customizing it. According to The Decoder, Ramp uses it to turn photos of receipts into structured data and expects ten times faster responses, and Scale AI adapted it for a vendor credit agent at Skydance with costs expected to drop to one-fortieth.

Two smaller Bolt models came alongside: a 74B model squeezed into 24GB of memory with 2-bit weights, and a 4GB version for voice assistants that follows later this month. The Decoder notes Super VL still trails the leading closed models by a wide margin on the overall Artificial Analysis ranking, and that everything is optimized for Nvidia hardware.

My take

Receipts, invoices, intake forms, signed PDFs: getting those into a CRM or accounting system as clean fields is where a lot of small business admin time goes.

An open model that reads documents well matters for one group in particular: clients whose documents should not leave their own servers, like clinics, lenders and accountants.

For most small teams, though, hosting a 120B model is not the first step. What I would do:

• Collect 50 real documents from the client, messy ones included.
• Define the exact fields you need and what a correct extraction looks like.
• Test a hosted model first, and score it against that set.
• Only move to a self-hosted open model if privacy rules or volume demand it, and run the same test set on it.

The model is replaceable. The test set and the field definitions are the asset.