Nvidia opens Nemotron 3.5 Super VL: agent model understands images, documents and video
The open model has 120 billion parameters, of which only around 12 billion are active per token. Nvidia promises up to five times more throughput on a single GPU.
Photo: Coolcaesar / Wikimedia Commons, CC BY-SA 4.0 (cropped and resized)
Video: YouTube · NVIDIA Developer · youtube.com
Nvidia on Thursday released Nemotron 3.5 Super VL, an open multimodal model for AI agents. It processes text, images, documents and video and is designed to reason, follow instructions and use tools. As a mixture-of-experts model it has 120 billion parameters, with around 12 billion active per token. According to the specialist site MarkTechPost, the context window holds up to one million tokens; the model accepts video as MP4 clips of up to two minutes.
Nvidia cites up to five times higher agent throughput on a single GPU than open models of the same size class. On the CharXiv chart test, the model scores 80.8 per cent according to the company, compared with 63 per cent for the smaller predecessor Nemotron 3 Nano Omni. These are the manufacturer's own figures; independent comparisons are still pending.
The weights are available on Hugging Face and ModelScope, along with training recipes, datasets and reinforcement-learning environments. According to MarkTechPost, the compressed NVFP4 version runs on, for example, a single DGX Spark or one H200 card. The model is also available via OpenRouter and the cloud services of Amazon, Google, Microsoft and Oracle.