NVIDIA Nemotron 3.5 Super VL lands in Amazon SageMaker JumpStart

NVIDIA Nemotron 3.5 Super VL is the vision-language member of the Nemotron 3.5 family, built for agentic workloads that must read and see. Its language backbone interleaves Mamba-2 and Mixture-of-Experts layers, initialized from Nemotron 3 Super Base and distilled from Nemotron 3 Ultra Base, while a C-RADIO vision encoder maps images and sampled video frames into the language model's embedding space. Training moved from continued pre-training through multimodal supervised fine-tuning, reinforcement learning with verifiable rewards and human feedback, multi-teacher on-policy distillation, and a final stage for the Multi-Token Prediction head.

The two JumpStart variants serve different needs. NVFP4 runs on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs and a checkpoint of about 75 GB, while BF16 requires ml.p5en.48xlarge with NVIDIA H200 GPUs and about 249 GB. NVIDIA's model cards report that NVFP4 matches BF16 on MMLongBench-Doc and stays within about one point on other multimodal benchmarks, scoring slightly higher on OCRBench v2 and OSWorld-Verified. That narrow gap may make the lighter NVFP4 default attractive for inference-heavy deployments where GPU footprint matters.

Inference performance measured on JumpStart showed mean output throughput of 63.07 tokens per second per request on ml.g7.24xlarge with NVFP4 and 140.85 tokens per second on ml.p5en.48xlarge with BF16, though AWS notes these run on different GPU systems and are not an isolated precision comparison. Time to first token stayed around half a second across configurations, but longer prompts change the picture: with the summarize profile of 10,000 input tokens, time to first token on ml.g7.24xlarge is 12.62 seconds. Teams weighing deployment choices will need to match the variant and serving configuration to their prompt lengths and latency targets.