AWS says NVIDIA Nemotron 3.5 Super VL joins other Nemotron models available on AWS. The model has 120 billion total parameters and activates 12 billion parameters per token. It accepts text, images, documents, and video, while producing text outputs, and supports context lengths of up to 1 million tokens.

According to the post, NVIDIA publishes the model’s weights, training datasets, and post-training recipes. Nemotron 3.5 Super VL uses a hybrid architecture combining Mamba-2, mixture-of-experts, and attention layers with a RADIO vision encoder. It is available in BF16 and NVFP4 checkpoints. NVIDIA reports scores of 63.44 percent on Hermes Bench, 87.05 percent on LangChain agent evaluations, and 80.8 percent on the CharXiv chart-reasoning benchmark.

AWS describes a reference implementation for deploying the model on standard Amazon EKS, EKS Auto Mode, and SageMaker HyperPod EKS. The setup uses GPU nodes, Amazon ECR for container images, Amazon FSx for Lustre for model weights, and Elastic Fabric Adapter for communication between GPUs.

The post says NVIDIA Dynamo can serve the model in an aggregated configuration or with prefill and decode handled on separate nodes. For vision-language workloads, image and video encoding can also run on a separate node. vLLM is included in the listed deployment stack, but the source does not attribute these deployment modes to vLLM.

The architecture was validated on p6-b200.48xlarge and p5en.48xlarge instances. AWS recommends NCCL tests for cluster validation and NVIDIA’s aiperf tool for measuring time to first token, inter-token latency, and throughput at different concurrency levels.