NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, giving operations and finance teams a downloadable AI model they can adapt to read documents and work with business tools. It comes with model weights—the files needed to run it—and customization resources, not a ready-made office application.
NVIDIA’s visual-investigation example turns document evidence into a proposal for human approval: the assistant reads a dashboard and scanned records, checks a database and recommends a change. It uses synthetic company data, and NVIDIA selected the recording from successful runs—not a representative production trial.
Running the model and training it need different budgets
The distinction matters for teams commissioning a document assistant. NVIDIA’s published configurations separate running its memory-saving NVFP4 version from adapting the model through its finance training recipe:
Task | NVIDIA’s documented setup |
|---|---|
Run the existing model | One DGX Spark or one H200 GPU, among other configurations. The Spark quick-start handles one request sequence at a time. |
Adapt it with the finance recipe | An eight-GPU machine with at least 80 GB per GPU and about 800 GB of storage for the full workflow. Training alone can run on four H200s. |
These are specific published setups, not universal minimums for every deployment or training method. The purchasing implication is to budget adaptation separately from the ongoing service: a machine that can answer questions need not be able to run this training recipe.
The OpenMDW 1.1 licence grants use without a licence fee, subject to its terms; computing, integration and maintenance still cost money. Baseten offers managed customization, so a team need not own the training cluster. Its announcement does not provide an all-in project price.
The free endpoint excludes confidential documents
OpenRouter lists a free hosted version with account-dependent usage limits. Its notice says not to upload confidential information or personal data, and says use is logged for security and NVIDIA product improvement. That makes it a route for exploring non-confidential, non-personal samples—not live private business records.
Start with one document task
NVIDIA publishes document and visual benchmark results, but those scores do not establish accuracy on a company’s own scans and tables. Before paying for customization, compare the unmodified model with the current workflow on one defined task: check critical figures, whether answers use the right source evidence, and how much correction a person must do.
If the job is only to extract text, layout or chart values, a narrower document tool also deserves evaluation; RohitAI’s LightOnOCR-3 guide covers that choice. Readers focused on high-volume text work can see the earlier Nemotron Lightning analysis.
AI-assisted reporting and analysis based on public announcements and documentation checked on October 15, 2026. RohitAI has not tested the model or hosted services.