Nemotron 3.5 Super

Nemotron 3.5 Super is NVIDIA’s vision-language model for coding, reasoning, tool use, and document understanding. It has 120 billion total parameters, with 12 billion active per token, and combines a hybrid Mamba-2/Transformer mixture-of-experts backbone with a C-RADIO vision encoder. NVIDIA model card

• Vision: ask questions about images, charts, screenshots, and document pages.
• Tools: build applications that call functions and incorporate their results.
• Thinking: choose whether to generate a reasoning trace before the answer.
• Long context: the model supports up to 1 million tokens. Available memory and runtime settings determine the usable context; prompts, images, history, tool results, and generated tokens share that budget. Model design

The architecture also includes Multi-Token Prediction (MTP) for speculative decoding. NVIDIA publishes BF16 and NVFP4 checkpoints, along with a separate LiLiCorr drafter.

License

NVIDIA publishes Nemotron 3.5 Super VL under the OpenMDW License Agreement, version 1.1. See the publisher’s model card for training details and evaluation results.