From the source
Amazon made
NVIDIA's Nemotron 3.5 Super VL vision-language model available in SageMaker JumpStart, where it can be deployed from SageMaker Studio or the SageMaker Python SDK.
- The model pairs a 120B-parameter hybrid Mixture-of-Experts language backbone with 12B active parameters and a C-RADIO vision encoder, accepting images, video and text and generating text.
- It supports context lengths of up to 1 million tokens and a built-in Multi-Token Prediction head for speculative decoding, with reasoning configurable per request.
- NVIDIA built it for document intelligence, GUI understanding for computer-use agents, video question answering and multimodal agentic applications.
- JumpStart offers two variants: an NVFP4 checkpoint quantized from BF16 with quantization-aware distillation, and BF16 full-precision reference weights.
- NVFP4 runs on G7 instances with RTX PRO 4500 Blackwell GPUs and BF16 on P5en instances with H200 GPUs; the default NVFP4 instance is ml.g7.24xlarge.
- In Amazon's test, BF16 on ml.p5en.48xlarge reached 140.85 output tokens per second per request, with time to first token about half a second across all three configurations.






