# AWS adds NVIDIA Nemotron 3.5 Super VL to SageMaker JumpStart > AWS lists NVFP4 inference and BF16 customization options for NVIDIA’s vision-language model, which accepts images, video and text. Canonical URL: https://www.devobs.io/news/news-event-7c7fa9bd55c5cce44dfa3875/ By: Maya Chen Published: 2026-10-17T14:30:51.979Z Updated: 2026-10-17T14:30:51.979Z Event date: 2026-10-16 Section: AI AWS announced on October 16, 2026, that NVIDIA’s Nemotron 3.5 Super VL is available through Amazon SageMaker JumpStart. According to the [AWS release notice](https://aws.amazon.com/about-aws/whats-new/2026/01/nvidia-nemotron-3.5-super-vl-on-sagemaker-jumpstart/), the vision-language model accepts images, video and text and generates text. AWS describes its intended uses as document intelligence, GUI understanding, video question answering and multimodal agents. ## Two variants for different workflows AWS lists two JumpStart variants in the [announcement](https://aws.amazon.com/about-aws/whats-new/2026/01/nvidia-nemotron-3.5-super-vl-on-sagemaker-jumpstart/): - **NVFP4 for inference:** AWS describes this variant as optimized for inference on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. - **BF16 for customization:** AWS lists reference weights on **ml.p5en.48xlarge**, with customization uses including supervised fine-tuning, reinforcement learning, distillation and domain adaptation. AWS says customers can deploy the model through the SageMaker console’s JumpStart catalog or the SageMaker Python SDK. The notice does not specify regional availability or pricing and does not explicitly label the release as a preview or general availability. For engineering teams, the distinction makes workflow and hardware early evaluation decisions. Establish whether the immediate goal is serving the supplied model or adapting it, then assess the corresponding variant. Check regional access and deployment costs before committing to either hardware target. ## Long context is not a performance guarantee In its [model description](https://aws.amazon.com/about-aws/whats-new/2026/01/nvidia-nemotron-3.5-super-vl-on-sagemaker-jumpstart/), AWS describes a 120-billion-parameter hybrid mixture-of-experts language backbone with 12 billion active parameters, paired with a vision encoder. It lists context lengths of up to one million tokens, structured JSON output, tool calling and per-request reasoning configuration. AWS also describes a built-in Multi-Token Prediction head for speculative decoding, intended to accelerate text generation. Those specifications provide an evaluation checklist, not measured application results. Test document extraction, video questions and GUI interpretation separately, measuring correctness and latency on representative inputs. Evaluate answer quality and response times across long inputs rather than treating the context limit as a performance guarantee. For workflows using JSON output or tools, validate generated data against the expected schema and authorize downstream actions outside the model. Schema validation and action authorization should be separate checks: a correctly formatted request should not automatically be permitted to change another system. ## Source references -