# Nvidia Releases Fully Open Nemotron 3.5 Models, Led by 120B-Parameter Super VL

*2026-10-17 02:00 UTC · AI · 1 min read*

> Nvidia published three open Nemotron 3.5 models with weights, datasets and training recipes, aimed at always-on and on-device AI agents.

Nvidia has released three new Nemotron 3.5 models as fully open releases, publishing the weights, training datasets and training recipes on Hugging Face. The flagship, Nemotron 3.5 Super VL, has 120 billion parameters and is pitched at fast, accurate "always on" agents. Two smaller Bolt models follow, including the 4GB Bolt E4B, designed for conversational agents that can run on devices.

The release follows Nemotron 3.5 Lightning in August and is the first since Nvidia agreed last month to acquire Hugging Face. Earlier the same week, the company also made Run.ai open source.

Omdia analyst Alexander Harrowell notes that small models have grown considerably larger, which helps enterprises fine-tune and customize them. He adds that Nvidia is partly acting out of self-interest, supporting the ecosystem around its GPUs, and that the new architecture scales memory better and allows a larger context window, which matters for agentic AI.

## Sources

- [techtarget.com](https://techtarget.com/ai/news/366652514/Open-source-and-on-device-the-strategy-behind-Nvidias-new-AI-models)

---
TechTools News · https://techtools.cz/tools/news/article/2530-nvidia-releases-fully-open-nemotron-3-5-models-led-by-120b-parameter-super-vl
Česká verze: https://techtools.cz/tools/news/clanek/2604-nvidia-vydala-plne-otevrene-modely-nemotron-3-5-vede-je-120miliardovy-super-vl