Nemotron 3.5 Super VL

Nemotron 3.5 Super VL is an open model from NVIDIA that reads text, images and video and answers with text. The letters VL mean vision and language. According to the model card, it has 120 billion parameters in total, of which 12 billion work on each step, it handles up to one million tokens of context, and it is meant for programming, searching, tool-assisted work, questions about documents and understanding of images and video. The release is dated 2026-10-12. This card links the full-precision version; a compressed NVFP4 version is published next to it.

Reasoning can be switched off, set to medium or fully on, so you can trade answer quality for speed. The model can predict several words ahead at once, which speeds up generation. The card lists English, Spanish, French, German, Italian and Japanese as supported languages. In the authors' own table the model scores 89.5 on the GPQA-Diamond science test and 88.2 on LiveCodeBench v6 programming; these are NVIDIA's figures, and in the same table some other open models score higher on several lines.

Hardware is stated clearly: the full-precision weights run on four H100 cards or two GB200; the compressed NVFP4 version runs on a single DGX Spark, a single H200, two H100 or a single GB200. It is designed for NVIDIA graphics cards and is served with vLLM or SGLang.

The weights are public without any acceptance step. The license is OpenMDW-1.1: it allows using, changing and sharing the model without restrictions on how the answers are used; you only keep the license text and notices when you share it, and your rights end if you sue over patents or copyright on the model. The card says the model is ready for commercial and non-commercial use. We found no territorial restrictions in the license text. NVIDIA also publishes the training data and recipes for the Nemotron family.

Date of check: 2026-10-21.