Nvidia Nemotron 3.5 Super VL: open document AI on four H100s

Nvidia released Nemotron 3.5 Super VL, an open-weights vision-language model built for documents, screens and AI agents. The model card on Hugging Face lists 120 billion total parameters, 12 billion active, a context window of up to 1 million tokens, and a license that allows commercial use. Nvidia also shipped a smaller Bolt 74B model packed into 24GB for local machines. For a business that reads invoices, forms and PDFs all day, free weights sound like a cost cut. Do the math first.

What actually happened

From Nvidia's model card:

• What it reads. Text, images and video in; text out. Nvidia lists reasoning, coding, tool use, document understanding and video understanding as the target jobs.
• Architecture. A Mamba-2 and transformer hybrid mixture-of-experts with a C-RADIO vision encoder. Built-in multi-token prediction speeds up decoding. Reasoning can be switched off, set to medium, or on.
• Hardware. The full-precision BF16 weights are listed for 4× H100 or 2× GB200. An NVFP4 version targets DGX Spark and data-center GPUs. Nvidia's quick start serves it with vLLM behind an OpenAI-compatible endpoint.
• License. OpenMDW-1.1. Nvidia says it is ready for commercial use.
• Benchmarks. These are Nvidia's own numbers, and they are mixed. It scores 71.1 on SWE-bench Verified and 68.1 on English OCRBench v2. On many rows it trails Qwen3.8-Flash-Next and GLM-5.3-Flash, which Nvidia lists in the same table. One standout: a 30.0 hallucination rate on AA-Omniscience, far lower than several rivals.

Why it matters for your business

Open weights you can run on your own hardware matter for one reason most of all: the documents never leave your building. If you handle medical forms, legal files or customer financials, that can matter more than raw benchmark rank.

But "free" is the weights, not the system. Four H100s is real money, rented or bought, and someone has to keep vLLM running. For most small teams, an API is cheaper until volume is high or data rules force the issue. Our rule of thumb:

• Run a 100-document test on your real files through the model on a rented GPU box and through your current API. Compare accuracy and cost per document.
• Self-host only if data can't leave, or monthly volume makes GPU hours cheaper than tokens.
• Keep the interface standard. An OpenAI-compatible endpoint means you can swap between hosted and self-hosted without a rewrite.
• Read the license before you build on it, as with any open model.

Sources: Nvidia on Hugging Face, OpenMDW License.