AWS Adds NVIDIA Nemotron 3.5 Super VL: SageMaker JumpStart Users Can Deploy an AI Model That Reads Images and Video

What happened

AWS announced that NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart. JumpStart is the catalog of pretrained models in SageMaker, AWS's machine learning service. The model can be deployed in SageMaker Studio or with a few lines of SageMaker Python SDK code.

According to AWS, Super VL is the vision-language model in NVIDIA's Nemotron 3.5 family, meaning an AI model that understands images and text together. It accepts images, video and text and generates text.

The model uses a mixture-of-experts (MoE) architecture: it is made up of many "expert" sub-networks, of which only a portion is active at a time. It has 120B total parameters with only 12B active per token, paired with a C-RADIO vision encoder. A token is the basic unit in which a model processes text.

The model supports a context of up to 1 million tokens, the amount of content it can handle at once. It also includes a built-in Multi-Token Prediction (MTP) head, a component that predicts several tokens at once to speed up generation. Reasoning mode (the model thinks step by step before answering) can be switched on or off per request.

AWS says NVIDIA designed it for four kinds of use: document intelligence, GUI understanding for computer-use agents, video question answering, and multimodal agent applications. Here, computer-use agents means AI programs that can operate software automatically. AWS also lists supported languages as English, French, German, Italian, Japanese, Spanish and Chinese, with the license being OpenMDW License Agreement version 1.1.

Comparing the two versions

AWS states that JumpStart offers two versions. BF16 is the full-precision reference weights, storing model values in a 16-bit format, usable for later custom training. NVFP4 is produced from BF16 through quantization-aware distillation. Quantization means lowering numerical precision to shrink the model; NVFP4 runs with 4-bit weights and activations and is aimed at optimized inference, that is, running the model to produce answers quickly and efficiently.

Quality benchmarks and limits

Citing the NVIDIA model card (the developer's published description of the model), AWS says NVFP4 matches BF16 on the MMLongBench-Doc benchmark at 62.82 for both, and stays within about one point on other multimodal benchmarks. Scores (BF16 and NVFP4) are: OCRBench v2 English 68.12 and 68.34; CharXiv reasoning 80.82 and 79.88; Video-MME v2 32.85 and 32.54; OSWorld-Verified 51.14 and 52.30.

On limits, AWS explains the model accepts images rather than raw PDFs, so pages must first be converted to images; a 1,536 x 1,024 page takes roughly 1,536 tokens of context. The JumpStart configuration accepts only one image and one video per request, with a serving context of 520 to 15,600 tokens.

Although the checkpoint includes an MTP head, the JumpStart configuration does not enable speculative decoding, the technique that uses MTP to speed up output. The endpoint, the service address that receives requests after deployment, serves only the base weights and does not load LoRA adapters. LoRA is a lightweight add-on module for fine-tuning.

AWS further notes that the endpoint runs with network isolation and cannot fetch URLs itself, so images and video must be embedded directly in the request as base64 data URLs, a way of writing the file itself as text inside the request.

What it means for general readers and businesses

For teams already on AWS, this means they can try a large model that reads images and watches video within a managed service, for scenarios such as invoice extraction and understanding charts and long documents. This is AWS's release announcement; actual results still need to be validated on your own data.

AWS reminds users that deployment requires two things: AWS account permissions to create SageMaker AI endpoints, and sufficient instance quota. For new accounts the ml.g7 quota starts at zero and must first be raised through Service Quotas. Deployment on ml.g7.24xlarge takes about 10 minutes. AWS also recommends deleting the endpoint, endpoint configuration and model after use to avoid ongoing charges.

Browse the latest news in this topic

• AI news