DEDUPEDNvidia expanded its open Nemotron family with three new models built for AI agents, covering everything from data center servers down to laptops and edge devices.
The headliner is Nemotron 3.5 Super VL, a 120 billion parameter vision-language model with 12 billion parameters active at a time. It is a hybrid Mamba-Transformer mixture-of-experts design with a 1 million token context window, aimed at specialized, always-on agents that run long tasks and need to read documents and video. Nvidia says it delivers the fastest task completion and highest accuracy in its class.
Two smaller models target local hardware. Nemotron 3.5 Bolt 74B-A9B fits in 24GB of memory and balances intelligence and speed for local agent workloads. Bolt E4B fits in 4GB for fast conversational agents. Super VL and Bolt 74B are available now, and Bolt E4B is due on October 23. Popular local tooling already supports the new models, so they can be run on your own machine on day one.
All three are open, which keeps Nvidia in an unusual spot: the chip supplier to every lab is also one of the most active publishers of free models.
Our take
Nvidia gives models away to sell the hardware that runs them, and that makes it one of the most reliable sources of open, well-documented models in the West.
From the issue
China wants to export-control its own AI models, and that changes your stack Friday, October 16, 2026