Most office AI still lives somewhere else. You upload a contract, wait for a remote server to read it, and hope the file stays private along the way. NVIDIA wants to flip that habit. On October 15, 2026, the company released Nemotron 3.5 Super VL as an open model, posted it on Hugging Face the same day, and framed it plainly: this one is built for agents that stay on, work through steps on their own, and run close to where the work actually happens.
That local part matters. A model that can sit on a single workstation changes who controls the documents, the timing, and the cost. For a law office with client files, a clinic with intake forms, or a finance team with statements it cannot send out, local is not a technical preference. It is the whole point.
A Big Model That Only Uses Part of Itself at a Time
Nemotron 3.5 Super VL has 120 billion total parameters, but only about 12 billion are active for each token it processes. That is the mixture-of-experts design in practice. Instead of waking up the entire network for every word, the model routes work to the experts it needs for that step. You get the benefit of a large model’s knowledge without paying the full compute bill on every pass.
The VL stands for vision-language, and the range is wide. The model reads text, images, short video, and PDFs, with a reported context window of 1 million tokens. A million tokens is not a tidy headline number. In daily terms, it means a long report, a stack of scanned pages, and the notes that go with them can stay in the same conversation. The agent does not have to forget page 12 by the time it reaches page 80.
NVIDIA lists the target jobs clearly: instruction following, tool use, multistep business workflows, visual understanding, document processing, coding, and research. Read that list again and you see a pattern. None of those jobs is a single question and a single answer. They are chains – read this, check that, open the other file, draft the summary, flag what does not match. That is exactly what an always-on agent is supposed to do while a person gets on with other work.
What NVIDIA Claims on Speed
Performance claims need a source, and here the source is NVIDIA itself. The company says Nemotron 3.5 Super VL delivers up to 5x higher throughput than comparable open models in its class on a single GPU. NVIDIA also says it matches its largest reasoning model at about one-fifth the parameter count on specialized tasks. Those are company claims reported across the release coverage, not independent lab results, and they are worth treating that way – interesting, specific, and still to be tested in each buyer’s own setup.
Even with that caution, the direction is clear. If a smaller active footprint can do specialized work well, more teams can afford to run an agent all day instead of renting bursts of compute for big jobs. That opens the door for smaller firms, not just large enterprises with dedicated clusters.
Two Bolt Models for Smaller Machines
The Super VL release did not arrive alone. NVIDIA also made available Nemotron 3.5 Bolt 74B-A9B 24GB on the same day. It has 74 billion parameters with 9 billion active, and it uses native 2-bit compression to fit in 24 GB of memory. NVIDIA tested it on a GeForce RTX 5090 PC and on DGX Spark. In plain terms, that is a high-end desktop and a compact desktop AI system – not a data center row.
A third model is coming fast. Nemotron 3.5 Bolt E4B 4GB, built for edge and device use, is due October 23, 2026. At 4 GB, the conversation shifts from desk to device: a field tablet, a factory floor unit, a point-of-sale system that needs to understand a form or a photo without a steady cloud link.
Three sizes, one family, one idea. Put the right size in the right place, and the agent can stay with the task instead of waiting on a network.
That approach sits well next to other moves to make agents more practical. Teams training agents by practice, not just by reading static data, are getting fresh backing – see how Labelbox is raising funds to train AI agents by practice for one example. And in finance, where document work never stops, Ramp’s push into AI finance tools shows how much demand there is for software that can read, sort, and act on routine paperwork without adding headcount for every step.
Open Weights, Commercial Use, and Real Choice
The license matters as much as the benchmarks. Nemotron 3.5 is released under OpenMDW-1.1, which allows commercial use, per the release reports. Weights are available via Hugging Face, NIM, OpenRouter, and ModelScope. For a business, that means you can try the model, tune it on your own documents, and build it into a product without starting with a licensing negotiation.
It also means comparisons get honest, fast. Open weights can be run, measured, and stress-tested by anyone. If the throughput claims hold up on a legal team’s contracts or a retailer’s product photos, people will say so. If they do not hold up on a particular task, that will show up too. Open releases invite that kind of checking, and the field is better for it.
Practical Takeaways for Teams Considering a Local Agent
Start with documents you already handle every week. Invoices, intake PDFs, support tickets with screenshots, research packs – pick one flow with a clear start and finish. Run the Super VL model if you need the full multimodal range and the long context, or start with Bolt 24GB if your workstation memory is the limit. Keep a person reviewing the output for the first weeks, and log where the agent hesitates. Those logs tell you whether you need a bigger model, better instructions, or cleaner source files.
Pay attention to privacy wins you can measure. How many files no longer leave the building? How many hours of copy-and-paste work disappear? Those numbers persuade a manager faster than any benchmark chart.
NVIDIA released these models on October 15, and the timing feels deliberate. Agents are moving from demos to daily work, and daily work happens on desks, in back offices, and on devices that cannot always phone home. A 120 billion parameter open model that activates only a slice of itself, reads video and PDFs, and has two smaller Bolt siblings for tighter hardware gives teams a real option to keep that work local. That is a useful, concrete step – and for offices that have waited to try agents until they could keep control of their own files, it might be the step that finally makes sense.
Be First to Comment