NVIDIA and AWS Bring Nemotron 3.5 Super VL to SageMaker JumpStart with 1 Million Token Window
NVIDIA's new vision-language model arrives in AWS's managed catalog with a hybrid mixture-of-experts architecture and quantized variants for agents that need to see and read documents
NVIDIA and Amazon Web Services announced the availability of Nemotron 3.5 Super VL in Amazon SageMaker JumpStart, Amazon's managed model catalog. The announcement was made on AWS's official machine learning blog.
The model can be deployed in a few steps in SageMaker Studio or with a few lines of the SageMaker Python SDK. It is the vision-language model (VLM) in the Nemotron 3.5 family, accepting images, video, and text as input and generating text as output.
Hybrid Mixture-of-Experts Architecture
Super VL combines a 120-billion-parameter hybrid Mixture-of-Experts (MoE) language backbone, with 12 billion active parameters per token, with a dedicated C-RADIO vision encoder. The configuration enables context windows of up to 1 million tokens.
The model also includes an integrated Multi-Token Prediction (MTP) head, which enables speculative decoding—a technique that accelerates generation by predicting multiple tokens ahead. Reasoning can be configured individually per request.
Focused on Agents That See and Read
According to NVIDIA, Super VL was built for four main areas: document intelligence, graphical user interface (GUI) understanding for computer-use agents, video question answering, and agentic multimodal applications.
The model targets workloads that require vision-language agents to combine reasoning, coding, tool use, document understanding, and video comprehension in a single system.
Two Variants and Measured Performance
JumpStart offers two variants of the model. NVFP4, the catalog default, is quantized from BF16 with quantization-aware distillation for optimized inference. BF16 provides the reference weights at full precision.
In multimodal benchmarks released by NVIDIA, the NVFP4 variant matches BF16 on MMLongBench-Doc and is within about one point on the other tests, with slightly higher performance on OCRBench v2 and OSWorld-Verified.
AWS measured inference performance with the JumpStart launch container, using a generation profile with 250 input tokens, 1,500 output tokens, and five concurrent streaming requests. The BF16 variant, running on ml.p5en.48xlarge instances, achieved 140.85 output tokens per second per request.
Time to first token was about half a second across the three configurations tested. With longer prompts, the picture changes: in the summarization profile with 10,000 input tokens, time to first token on the ml.g7.24xlarge instance reaches 12.62 seconds.
Validated Deployment
The NVFP4 variant runs on G7 instances with NVIDIA RTX PRO 4500 Blackwell GPUs, offering a smaller GPU footprint. AWS notes that because the variants run on different GPU systems and configurations, the measurements do not constitute an isolated accuracy comparison between BF16 and NVFP4.
The launch expands access to the Nemotron 3.5 family within the AWS ecosystem, allowing developers to deploy the vision-language model without managing their own infrastructure. The full technical report, including comparisons against other public models and evaluation adjustments, was published by NVIDIA Research.