Key points
- Super VL is a 120B-parameter model that activates 12B per token; weights, datasets and training recipes are published, with enterprise deployment through Nvidia NIM.
- Nvidia says it matches the larger Nemotron 3 Ultra on agent and image tasks at about one-fifth the parameters.
- A 24 GB Bolt variant for local machines is out now; a 4 GB voice-focused version is due later this month.
- Early users named by Nvidia include Ramp, EY, Hebbia, CrowdStrike and Salesforce.
Why it matters for marketers: Teams building document-heavy or always-on agents get an open option they can fine-tune and host themselves, though the tuning for Nvidia hardware will shape the real cost.
Nvidia has put out Nemotron 3.5 Super VL, a large open model meant for agents that work in the background and read text, images, PDFs and video. Two smaller Bolt models aim at local machines, with one available now and a lighter voice-oriented version due later this month.
Nvidia says the model was distilled from its biggest reasoning model and keeps similar accuracy on agent and image tasks at a fraction of the size. Those figures come from the company, though it cites an independent test by EXO Labs on speed and accuracy. On document work, Nvidia says the model can read a chart, connect it to a table elsewhere in the file and summarize both without a separate vision model.
For marketing and content teams, the draw is control. Nvidia pitches the model to companies that want to fine-tune on their own data and run it on their own servers instead of relying on a provider's terms. Tasks like extracting data from receipts, reviewing contracts or processing large volumes of documents are the use cases it highlights. It also releases training data and recipes, which is more than many open-weight releases include.
There are caveats. The local Bolt model still needs 24 GB of memory, which limits it to developers and powerful workstations for now. Nvidia tunes the whole Nemotron family for its own hardware, so costs and performance will depend on that stack. Teams should test the model on their own documents and workflows, and watch for the smaller Bolt release before judging whether on-device agents are practical.
Sources
This brief is our own summary. Read the full reporting at the source.
