NVIDIA Launches Open Nemotron 3.5 Models for Agents
NVIDIA released three new open-weight AI models on October 15, 2026, and they are all built for one job: running AI agents. The lineup is led by Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise agents, plus two smaller Bolt models designed to run on local hardware.
This is NVIDIA making its biggest open-model push yet into agentic AI. And the timing is not random.
The three models
Nemotron 3.5 Super VL is the flagship. 120 billion parameters total, with about 12 billion active per token thanks to a mixture-of-experts design. It reads text, images, documents, charts, and video in a single model. No need to bolt a vision model onto a text model. One agent can read a chart, cross-check a table later in the same report, and summarize both.
NVIDIA claims it is the fastest open agentic model in its size class, with up to 5x higher throughput than rivals on agent tasks. It handles instruction following, tool use, and multistep business workflows. Context window: 1 million tokens. It is built for always-on agents that run for days or weeks doing jobs like document review, web research, and cybersecurity triage.
Nemotron 3.5 Bolt 74B-A9B squeezes 74 billion parameters into 24GB of memory. That fits on a single high-end workstation, NVIDIA tested it on a GeForce RTX 5090. It targets local agent workloads where companies want power without data-center bills.
Nemotron 3.5 Bolt E4B is the tiny one. 4GB footprint for fast conversational agents on edge devices. It arrives October 23, while the other two are available now.
Why open matters here
NVIDIA is releasing the weights, the datasets, and the training recipes. Companies can post-train these models on their own proprietary data, tools, and workflows, then deploy on their own infrastructure. For enterprises worried about sending sensitive documents to someone else’s API, that is the whole pitch. Your data never leaves your building.
The license is NVIDIA’s OpenMDW-1.1, which allows commercial use, fine-tuning, and redistribution. That puts these models in direct competition with open offerings from Meta, Alibaba, and the growing field of agent-focused releases.
The bigger picture
Step back and the strategy is clear. NVIDIA’s AI hardware business is the engine of the AI boom, but hardware alone does not lock in developers. Open models that run best on NVIDIA chips do. Every company that post-trains Nemotron and deploys it on NVIDIA GPUs is another customer for the hardware stack.
It also lands as AI infrastructure spending faces harder questions about returns. Open, efficient models that enterprises can run themselves are NVIDIA’s answer to the cost debate. Why rent someone else’s giant model when you can run a specialized open one on hardware you already own?
The performance claims deserve the usual caution. “5x faster” and “leading accuracy” are NVIDIA’s numbers on NVIDIA’s chosen benchmarks. Independent testing will tell the real story. But the direction is unmistakable. The agent era needs open models, and NVIDIA wants to own that layer too.
FAQs
What are the three NVIDIA Nemotron 3.5 models?
Super VL (120B parameters, 12B active) for enterprise multimodal agents, Bolt 74B-A9B (fits 24GB) for local agent workloads, and Bolt E4B (4GB) for fast conversational agents on edge devices.
Are the models free to use commercially?
The weights are open under NVIDIA’s OpenMDW-1.1 license, which allows commercial use, fine-tuning, and redistribution. Enterprise production deployment through NVIDIA AI Enterprise may carry separate licensing costs.
When are the models available?
Super VL and Bolt 74B-A9B are available now (released October 15, 2026). Bolt E4B arrives October 23, 2026.
