If you run a clinic, a small law firm, or a bookkeeping practice, you have probably had this conversation already. Someone pitches an AI agent, you ask where the client files go, they say “our cloud,” and that is the end of the meeting. Nvidia’s latest release gives you a reason to keep talking a little longer.
What shipped
Nvidia announced a new set of open Nemotron 3.5 models on October 15, and TechCrunch covered them on October 19. There are three.
Nemotron 3.5 Super VL is a 120-billion-parameter multimodal model built for enterprise agents. Nvidia says it matches its largest reasoning model on agent tasks at about a fifth of the size, and claims it is five times faster than the next-best open model in its class. Those are Nvidia’s own numbers, and no independent test has confirmed them yet.
Nemotron 3.5 Bolt 74B-A9B 24GB has 74 billion parameters, 9 billion of them active, and uses 2-bit quantization to fit in 24GB of memory. Nvidia built it to run agents locally on its new RTX Spark laptops and DGX Spark workstations. The model card lists text, image and video input, with up to 256,000 tokens of context.
A smaller model for conversational agents, Bolt E4B 4GB, is due October 23 according to Nvidia. Super VL and Bolt 74B are already on Hugging Face, and Nvidia’s blog lists support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
Why this matters for a business that cannot move its data
Open models tuned for agent work are not new. What is different here is a model sized to run on a machine that fits under a desk, tuned for the tool calls, browsing and document reading an agent actually does.
That matters most where the data itself is the risk: patient records, privileged documents, client financials, anything a compliance officer worries about. An agent could read scanned intake forms or work through a discovery folder on hardware you own, behind your own firewall, and Nvidia’s pitch that the data never leaves the building describes a real use in that setting.
Where to be careful
Local makes sense when the data is sensitive and the work is repetitive, with answers a person can check. Nvidia itself describes Bolt as “smart enough to get most jobs done,” which fits that kind of structured work well. Stick with an API when a task needs the strongest reasoning available, when the data is not sensitive, or when nobody on staff can keep a machine running. Check the rest of the workflow too. If half of it still calls out to cloud services, running the model locally does not make the whole thing private.
The hardware is a real cost. The Verge reports RTX Spark laptops start at $2,599 with 24GB of unified memory and go up to $6,999.99 for 128GB. A 24GB model on a 24GB machine leaves nothing else to run, so budget for more memory than the model strictly needs, and treat the machine as a dedicated agent box, not the one someone also uses for email.
Bolt ships under the OpenMDW-1.1 license, which allows commercial use and modification, puts no restrictions on outputs, and comes with no warranty. You are responsible for clearing any rights and consents, which is standard for open weights and means you own the consequences when the agent gets something wrong. Nvidia’s own model card also warns that the model can be swayed by untrusted input such as prompt injection and should run with guardrails. Any agent that opens email attachments is exposed to exactly that.
What still takes the work
Running the model locally is now the easy step. Whether the project actually works depends on everything around it: which folders the agent can read, which actions need a person’s sign-off, what gets logged, who updates the model, and who notices when a vendor changes a form layout and the extraction starts failing silently.
If you have held off on agents because your data cannot leave the office, pick one workflow and get the permissions right before you add a second.
Sources
- Nvidia launches new open models to run agents in the enterprise and on PCs — TechCrunch, October 19, 2026
- Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models — NVIDIA Blog
- NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB model card — Hugging Face
- The first Nvidia RTX Spark laptops cost up to $7,000 — The Verge
Kush AI Automation builds in accounts you own, and nothing that touches sensitive records runs without the permissions and logging to show for it. Bring the workflow your compliance person worries about most and book a free call.
