Nemotron 3.5 is NVIDIA's open model family for agents, released on October 15, 2026. It is really three models. Super VL is the flagship: 120 billion parameters in a mixture-of-experts design (about 12 billion active per token), multimodal, with a 1 million token context, built to read text, images, short video and PDFs and to run multi-step business workflows. Bolt 74B-A9B squeezes a 74 billion parameter model into 24GB of memory with 2-bit quantization, so it runs on a single high-end workstation. Bolt E4B, due October 23, is a 4GB model for edge devices and fast voice agents.

All three ship as open weights under the OpenMDW-1.1 license, which allows commercial use. That changes the conversation for a lot of teams: the model can live on your hardware, be post-trained on your documents, and never send client data to anyone's cloud. NVIDIA's speed and accuracy claims are its own; the use cases below follow its launch material and first-day coverage, and the interesting ones are the workflows that were previously awkward or off-limits.

Use cases for operations and office teams (Super VL)

  • Document intelligence over PDFs. Super VL reads tables, charts and page layouts, not just raw text. Point it at a folder of invoices, inspection reports or supplier quotes and get structured rows out, with the chart on page 3 reconciled against the table on page 20.
  • Always-on research agents. NVIDIA pitches Super VL for jobs that run for hours or days: competitor monitoring, supplier research, evidence gathering across many websites, with answers verified against sources rather than answered from memory.
  • Customer service agents that follow the actual policy. Ground the agent in your returns policy and product manuals and let it handle multi-step cases with tool calls, instead of a chatbot that improvises. Because it is open, you can post-train it on your resolved tickets.
  • Video review and search. Super VL accepts short video input, so meeting recordings, training videos and site footage can be summarized and made searchable instead of sitting unwatched in a drive.

Use cases for developers

  • A coding agent that can see. Early integrations such as Kilo pair Super VL with an IDE agent: paste a screenshot of a broken layout and the agent works from what it sees, then edits code, runs tests and reviews the diff.
  • Cybersecurity triage. Sort alerts and logs, group duplicates, and rank what a human should look at first. NVIDIA lists security triage among Super VL's target agent jobs.
  • A specialist you train yourself. Post-train Super VL on your own data, tools and workflows and deploy it on your infrastructure. NVIDIA says a post-trained Super VL reaches accuracy comparable to its largest reasoning model at about a fifth of the parameter count; treat that as a vendor claim to test, not a given.

Use cases for privacy-sensitive teams (Bolt 74B, local)

  • A client-document assistant that never phones home. A law firm, clinic or accounting practice can run Bolt 74B on one RTX workstation and let staff query contracts, files and records locally. The documents never leave the building, which is the whole point for regulated work.
  • A personal research and writing agent. Bolt is designed to sit resident on a personal machine and take delegated work through local harnesses such as Ollama, LM Studio or llama.cpp: research a purchase, draft the summary, debug the script, with no cloud bill and no uploads.

Use cases at the edge (Bolt E4B, from October 23)

  • Fast local voice and device assistants. The 4GB E4B model targets GeForce RTX PCs, Jetson boards and embedded interfaces: kiosks, in-store helpers, and voice control that responds immediately because nothing round-trips to a data center.

Access and pricing

The weights are free on Hugging Face under the OpenMDW-1.1 license (commercial use allowed), and Super VL is also packaged as an NVIDIA NIM microservice. Your real costs are hardware and time: a Bolt-capable workstation, or server GPUs for Super VL, or cloud GPU rental. NVIDIA sells supported enterprise deployment through its AI Enterprise program; trade coverage reports pricing around $4,500 per GPU per year, which is optional for self-hosters. Bolt E4B arrives October 23, 2026.

Frequently asked questions

Is NVIDIA Nemotron 3.5 available yet?
Mostly yes. Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B were released on October 15, 2026, with weights on Hugging Face the same day. The smallest model, Bolt E4B, follows on October 23, 2026.
Is Nemotron 3.5 free?
The model weights are free to download and use commercially under NVIDIA's OpenMDW-1.1 license. Running them still costs hardware or cloud time, and NVIDIA sells supported enterprise deployment separately; trade coverage reports an AI Enterprise license in the region of $4,500 per GPU per year, which most small teams will not need to self-host.
What hardware do I need to run Nemotron 3.5?
It depends on the model. Bolt 74B-A9B is quantized to fit in 24GB of memory, so a workstation with an RTX 5090 or a DGX Spark can run it. Super VL is a server-class model: NVIDIA lists a DGX Spark, one H200, two H100s or a GB200. Bolt E4B targets 4GB edge devices such as Jetson boards.
What can Nemotron 3.5 do that a chat model cannot?
Super VL was built for agents, not chat: it reads text, images, short video and PDFs in one model, follows instructions across multi-step tasks, and calls tools. The practical difference is document work: it can read a chart on page 3, match it with a table on page 20, and summarize both.
Can I fine-tune Nemotron 3.5 on my company's data?
Yes, and that is the point of the open release. NVIDIA ships weights, datasets and training recipes, and organizations can post-train Super VL on their own documents, tools and workflows, then deploy it on their own infrastructure.

Sources

More from this hub: the October 2026 launch wave hub, Gemini 4 Argon use cases and Nano Banana 2.1 use cases. Thinking about local AI for your team? Book a workshop and we will work out the hardware and workflow math with you.