AWS announced on its Machine Learning Blog that NVIDIA Nemotron 3.5 Super VL can now be deployed from Amazon SageMaker JumpStart, either through SageMaker Studio or a few lines of the SageMaker Python SDK.
What the model is
Super VL is the vision-language member of the Nemotron 3.5 family. It takes images, video and text as input and produces text. It has 120B parameters in total, of which 12B are active per token, and pairs a hybrid Mixture-of-Experts language backbone with a C-RADIO vision encoder. The model supports context of up to 1 million tokens, and reasoning can be switched on or off per request. It also supports tool calling and JSON schema output. Listed languages are English, French, German, Italian, Japanese, Spanish and Chinese. The license is the OpenMDW License Agreement, version 1.1.
AWS points to four uses: document intelligence (OCR, charts, tables, multipage files), GUI understanding for agents that operate applications, short video question answering, and multimodal agents.
Two variants
- NVFP4: a quantized version built for lower latency and higher throughput. It is about 75 GB and runs by default on ml.g7.24xlarge, with ml.g7.48xlarge as the larger option.
- BF16: the full-precision reference weights, about 249 GB, on ml.p5en.48xlarge. They are meant for customization such as fine-tuning, distillation or building your own quantization. The JumpStart endpoint itself serves base weights only and does not load LoRA adapters.
In the NVIDIA model-card results quoted by AWS, NVFP4 matches BF16 on MMLongBench-Doc (62.82) and stays within about one point on the other multimodal tests. It scores 68.34 against 68.12 on OCRBench v2 and 52.30 against 51.14 on OSWorld-Verified.
Measured speed
The benchmark used 250 input tokens and 1,500 output tokens, with five streaming requests running at the same time. NVFP4 on ml.g7.24xlarge produced 63.07 tokens per second per request, with 0.45 seconds to first token. BF16 on ml.p5en.48xlarge produced 140.85 tokens per second with 0.53 seconds. AWS notes that the two variants ran on different hardware, so this is not a clean precision comparison. With a 10,000-token prompt, time to first token on ml.g7.24xlarge rose to 12.62 seconds.
Limits to know before building
- The JumpStart configurations serve 520 to 15,600 tokens of context, far below the model's 1 million tokens.
- One image and one video per request. A page image of 1,536 x 1,024 costs about 1,536 tokens.
- The model takes images, not raw PDFs, so each page must be rendered first.
- Media must be sent inline as base64, because the endpoint runs with network isolation.
- Speculative decoding is not enabled in the JumpStart configuration, even though the checkpoint includes the head for it.
- New accounts begin with a zero quota for ml.g7, so submit a request for a higher limit prior to rolling out your deployment. Deployment takes about 10 minutes on ml.g7.24xlarge.
The endpoint exposes a vLLM-based chat completions API, so existing chat-style integrations need little change.