NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart

NVIDIA has made the Nemotron 3.5 Super VL model available in Amazon SageMaker JumpStart. This vision-language model can process images, video, and text, generating text in return. It combines a 120B-parameter hybrid Mixture-of-Experts language backbone with a 12B active parameter C-RADIO vision encoder. The model supports context lengths up to 1 million tokens and includes a Multi-Token Prediction head for speculative decoding.

Two variants are offered in JumpStart: NVFP4, which is quantized from BF16 with quantization-aware distillation for optimized inference, and the BF16 full-precision reference weights. Both variants support English, French, German, Italian, Japanese, Spanish, and Chinese. The NVFP4 variant is intended for optimized inference, while the BF16 variant allows for customization through further training or domain adaptation.

NVIDIA built Super VL for document intelligence, GUI understanding, video question answering, and multimodal agentic applications. The model was trained using continued pre-training, long-context and software-engineering mid-training, vision-language training, and a two-stage multimodal supervised fine-tuning process. RL with verifiable rewards and human feedback, multi-teacher on-policy distillation, and a final stage further training the MTP head were also involved in the training process.

The NVFP4 checkpoint is produced from BF16 with NVIDIA Model Optimizer and quantization-aware distillation, running with 4-bit weights and activations on Blackwell GPUs. The JumpStart endpoint serves the base weights only, without loading LoRA adapters. The model is licensed under the OpenMDW License Agreement, version 1.1.