NVIDIA Nemotron 3.5 Super VL Debuts on Amazon SageMaker JumpStart with Advanced Vision-Language Capabilities

NVIDIA’s Nemotron 3.5 Super VL, a multimodal vision-language model (VLM), is now accessible through Amazon SageMaker JumpStart, enabling developers to deploy a powerful AI agent capable of processing images, video, and text inputs with minimal setup. This addition broadens SageMaker’s AI portfolio, focusing on tasks that require integrated visual and textual understanding.

Super VL incorporates a 120-billion parameter hybrid Mixture-of-Experts (MoE) architecture with 12 billion active parameters per token. It features a dedicated C-RADIO vision encoder alongside a language backbone composed of layers that interleave specialized Mamba-2 and MoE components, optimized for extended context handling up to one million tokens.

The model supports multiple use cases, including document intelligence—such as extracting information from images of multi-page documents or charts—GUI understanding for computer-use agents, video question answering, and multimodal agentic interactions involving reasoning and tool use. JumpStart provides two variants: NVFP4, a quantized model optimized for inference on NVIDIA RTX PRO 4500 Blackwell GPUs, and a BF16 full-precision variant designed for high fidelity on NVIDIA H200 GPUs.

Training involved several stages, from pre-training to supervised fine-tuning with multimodal data, reinforcement learning with human feedback, and multi-teacher distillation. The NVFP4 model is produced using quantization-aware distillation to reduce computational demand without substantial loss in accuracy, making it suitable for deployment on cost-effective blackwell architecture instances.

In real-world benchmark tests within SageMaker, NVFP4 closely matches the full-precision BF16 variant in accuracy on multimodal reasoning tasks while requiring fewer computational resources. The BF16 variant achieved output speeds of over 140 tokens per second on high-end ml.p5en.48xlarge instances, demonstrating robust performance for demanding applications.

Deploying Nemotron 3.5 Super VL is streamlined across SageMaker Studio, the SageMaker Python SDK, and Hugging Face’s model hub integration, with configurable inference profiles supporting various context lengths and reasoning modes. Media inputs must be preprocessed accordingly, such as rendering PDFs to images due to the model’s image-based input requirement.

While Nemotron 3.5 Super VL significantly advances multimodal AI capabilities on AWS infrastructure, practical limitations include request payload size constraints, the necessity of media pre-processing, and current support for single-image or single-video-input requests per call. Speculative decoding is included but not enabled in the default serving configuration, leaving room for future performance tuning.

Overall, integration of NVIDIA’s Nemotron 3.5 Super VL into Amazon SageMaker JumpStart marks a meaningful enhancement in the accessibility of high-capacity VLMs, facilitating advanced AI functionalities for enterprises and developers tackling complex multimodal understanding tasks.