Nvidia launches Nemotron 3.5 Super VL, an open 120-billion-parameter model

Nvidia's Nemotron 3.5 launch puts it in the open-model arena alongside other frontier labs, but with a specific pitch: an open 120-billion-parameter model that the company says rivals its own largest reasoning systems at one-fifth the size. By releasing the Super VL weights openly, Nvidia is targeting organizations that want to post-train on proprietary data and keep the model behind their firewall — a sovereignty and privacy argument that closed frontier models cannot match.

The company trained Super VL with open data and open methods using about 26 trillion tokens, 3.8 million post-training samples and reinforcement learning across 100 open environments. It is also releasing open datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, safety, fairness and vision-language understanding, which could make it easier for other developers to build on its work.

The Bolt models extend the same family to local hardware. Bolt 74B-A9B fits within 24 gigabytes of memory, letting users delegate everyday tasks like writing and debugging code on an enterprise workstation or PC GPUs. A smaller Bolt E4B 4GB due later this month targets edge devices such as wearables, car dashboards, point-of-service terminals and conversational interfaces.