AWS brings NVIDIA’s Nemotron 3.5 Super VL to SageMaker JumpStart

NVIDIA’s Nemotron 3.5 Super VL is now available through Amazon SageMaker JumpStart, giving AWS customers a deployable vision-language model for documents, video and computer-use agents. The practical choice is between a quantised version for inference on smaller GPU instances and full-precision weights for customisation, though the advertised million-token context is not available in the listed JumpStart serving configurations.

AWS AI Watch analysis

What happened

AWS says the 120-billion-parameter model has 12 billion active parameters per token and accepts images, video and text, generating text in response. JumpStart offers an NVFP4 variant, aimed at optimised inference, and BF16 reference weights for custom training and adaptation. The NVFP4 checkpoint is about 75 GB; BF16 is about 249 GB.

The model is designed for tasks including document understanding, chart and OCR work, video questions and GUI understanding for computer-use agents. AWS’s serving configurations offer context lengths from 520 to 15,600 tokens, rather than the model’s maximum of one million. The AWS announcement and deployment details include setup instructions and AWS’s own performance measurements.

Our top picks

• Two deployment options
  NVFP4 runs on G7 instances; BF16 reference weights run on an H200-based P5en instance.
• One model for images and video
  The model takes images, video and text as input, targeting document, screen and video understanding.
• A useful deployment trade-off
  AWS’s NVFP4 option uses a smaller GPU footprint than its BF16 setup, while retaining broadly similar results on the selected multimodal benchmarks.
• Customisation, with a boundary
  The BF16 weights support fine-tuning and adaptation, but the JumpStart endpoint does not load LoRA adapters.
• Performance figures with a caveat
  AWS measured 63.07 output tokens per second per request on its default NVFP4 instance and 140.85 on the BF16 setup. Those runs use different GPU systems, so they are not a clean precision-versus-precision comparison.

Why it matters

JumpStart gives AWS users a managed route to try a sizeable multimodal model without building the serving setup from scratch. The two variants make the trade-off clearer: a smaller-footprint inference option, or heavier reference weights intended for teams that want to customise the model.

The headline context window needs a little translation before anyone orders GPUs. AWS says the JumpStart configurations serve up to 15,600 tokens, and its performance test found that a 10,000-token input pushed first-token latency on the default NVFP4 instance to 12.62 seconds. Workload and serving configuration matter; the million-token headline is not the endpoint’s default experience.

Our read

This is a useful addition for AWS teams with document, video or screen-based workloads, and the paired checkpoints offer a practical starting point for testing. Treat the benchmark table as AWS’s deployment data, not a universal ranking: compare the model on your own prompts, latency needs and GPU budget before making it a production choice. A million-token context window is a fine headline; the configuration you can actually deploy is the more useful number.

What to watch

• Whether AWS adds longer-context JumpStart serving configurations.
• How NVFP4 and BF16 perform on the same hardware and workload.
• Whether teams can customise the deployed service beyond the listed reference-weight options.

Discussion spark: For multimodal agent work, would you prioritise a smaller-footprint model that is easier to serve, or full-precision weights you can customise?

not affiliated with or endorsed by Amazon Web Services (AWS)