NVIDIA's Open Nemotron 3.5 Models: Run AI Agents Inside Your Company — Your Data Never Leaves
What NVIDIA announced
On October 15, 2026, NVIDIA released three open-weight models in the Nemotron 3.5 family — downloadable, runnable on your own hardware.
Nemotron 3.5 Super VL: 120B-parameter mixture-of-experts (12B active per token), multimodal — reads text, images, PDFs, tables, charts, short video. For always-on enterprise agents; available now.
Nemotron 3.5 Bolt 74B-A9B 24GB: compressed to fit in 24 GB of memory — a single desktop workstation with a strong RTX GPU. Available now.
Nemotron 3.5 Bolt E4B 4GB: a lightweight model for conversational agents, arriving October 23.
They ship under the OpenMDW-1.1 license — commercial use allowed — with weights, datasets, and training recipes. You don't rent intelligence; you own it.
Why this matters to your business
AI has mostly been a cloud service — your data on someone else's servers, a monthly bill, and trust in their privacy. Three problems follow:
First: sensitive data. Contracts, pricing, payroll — all through a third party. For legal, healthcare, and finance, that's a deal-breaker.
Second: rising costs. Gartner forecasts agentic-workflow inference costs rising fivefold by 2028. Deeper dependence, bigger bill.
Third: reliability. Outages, policy changes, sudden price hikes — all outside your control.
Local open models flip the equation: intelligence lives inside your company, data never leaves, and costs flatten after setup.
What "open" means commercially
You're the owner, not the tenant. A closed model dies when your subscription does. An open model keeps working even if NVIDIA changes its policies tomorrow.
Train it on your data. Published training recipes let you fine-tune the model on your policies, products, and service style — an agent that speaks your company's language.
Full privacy. The agent reading your invoices runs on a server in your office — nothing leaves your premises. Local AI automation becomes a genuine option for industries that rejected the cloud outright.
4 practical uses inside your company
1. Document-reading agents. Super VL reads tables, charts, and scanned PDFs — extracting invoice data, matching purchase orders, flagging mismatches, unseen by any human.
2. Contract review. It reads full contracts, extracts key clauses, and checks internal policies. Days become minutes — inside your walls.
3. An internal assistant. Fine-tuned on your handbooks, it answers recurring questions — leave policy, quoting procedures — without interrupting management.
4. Chatbots that actually act. An AI chatbot on a local model that understands context and completes tasks: quotes, bookings, order tracking — all inside your infrastructure.
This extends our piece on AI agents as digital employees — now inside your walls.
Before you go local: 3 questions
1. Do you have the hardware? The large model needs a GPU server; the mid-size one runs on an RTX desktop. One-time cost vs. annual cloud bill — local usually wins from year one.
2. Is your data ready? Its real power appears after fine-tuning on your data. Scattered documents? Organize first.
3. Who builds and maintains it? Local deployment needs real expertise — professional AI integration covers architecture, model selection, and proper agent builds.
Unsure? Start with a technical feasibility study — numbers beat headlines.
The bottom line
NVIDIA's open Nemotron 3.5 release declares that powerful AI is no longer cloud-only. The question is no longer "should we use AI?" but "where does our AI live?"
At Harisl we build software solutions and AI agents inside your infrastructure — model selection, training on your data, ongoing maintenance. If "AI that never leaves your company" interests you, contact us and let's map the road together.