NVIDIA Nemotron 3.5 Super VL arrives on SageMaker with 1M token context

Amazon Web Services announced on November 9, 2026, that NVIDIA Nemotron 3.5 Super VL is available for deployment via Amazon SageMaker JumpStart. This release provides engineers with immediate access to a high-capacity vision-language model designed for complex multimodal tasks, including document intelligence and video understanding.

What happened

The new model joins the NVIDIA Nemotron 3.5 family as its dedicated vision-language component. It accepts images, video, and text as inputs and generates text responses. The architecture combines a massive 120-billion-parameter hybrid Mixture-of-Experts language backbone with a specialized C-RADIO vision encoder. Despite its total size, the model activates only 12 billion parameters per token, balancing computational load with performance.

SageMaker JumpStart offers two distinct variants to suit different engineering needs. The default option is NVFP4, a quantized version optimized for latency and throughput on newer GPU hardware. The second option provides BF16 full-precision reference weights, intended for teams that plan to perform further customization, such as supervised fine-tuning or domain adaptation. Both variants support context lengths of up to one million tokens, enabling the processing of extensive documents or long video sequences.

The deployment process is streamlined through SageMaker Studio or the SageMaker Python SDK. Engineers can select the model card, choose their preferred instance type, and deploy the endpoint in approximately ten minutes. The service exposes a vLLM-powered chat completions API, allowing developers to integrate the model using familiar request formats while handling media inputs as base64-encoded data URLs.

How it works

Nemotron 3.5 Super VL uses a hybrid architecture that interleaves Mamba-2 and Mixture-of-Experts layers within its 88-layer language backbone. This design allows the model to route tasks to specific experts, activating only a subset of its 512 available experts for any given token. A shared Multi-Token Prediction head supports speculative decoding, although this feature is not enabled in the default JumpStart serving configuration. The vision encoder processes images and sampled video frames, mapping them into the language backbone’s embedding space via a vision-language connector.

For inference, the model distinguishes between reasoning-heavy tasks and direct extraction. Reasoning is enabled by default, allowing the model to think through complex problems before answering. Developers can disable this feature for simpler tasks like optical character recognition or JSON extraction to reduce latency. The NVFP4 variant uses 4-bit weights and activations on Blackwell GPUs, keeping the vision components in higher precision to maintain accuracy while reducing memory footprint.

Key details

• Architecture: 120B total parameters with 12B active per token, using a hybrid Mamba-2 and MoE backbone.
• Context window: Supports up to 1 million tokens, with serving configurations ranging from 520 to 15,600 tokens.
• Variants: NVFP4 for optimized inference on G7 instances and BF16 for customization on P5en instances.
• Inputs: Accepts text, images, and short MP4 video clips, with media provided as base64-encoded data URLs.
• Languages: Supports English, French, German, Italian, Japanese, Spanish, and Chinese.
• License: Released under the OpenMDW License Agreement, version 1.1.

Why it matters

For software builders, the availability of this model on SageMaker JumpStart lowers the barrier to entry for advanced multimodal applications. The ability to process long contexts means engineers can build agents that analyze entire technical manuals, legal contracts, or lengthy video logs without splitting them into arbitrary chunks. This reduces the complexity of application logic and improves the coherence of the model’s responses across large datasets.

The distinction between the NVFP4 and BF16 variants also offers strategic flexibility. Teams focused on production inference can leverage the quantized NVFP4 model on cost-effective G7 instances to achieve low latency. Meanwhile, research and development teams can use the BF16 weights to fine-tune the model for specific industry verticals, such as healthcare or finance, before deploying a customized version. This dual approach supports both rapid prototyping and deep specialization.

What you can do

• Request service quota increases for ml.g7.24xlarge or ml.p5en.48xlarge instances in your AWS account before attempting deployment.
• Use the SageMaker Python SDK v3.25 or later to automate the deployment process and manage endpoint configurations programmatically.
• Render PDF pages to PNG images before sending them to the model, as the endpoint does not accept raw PDF files directly.
• Disable reasoning via enable_thinking=False for straightforward extraction tasks to improve response speed and reduce token usage.
• Set max_tokens to at least 8,192 when performing video analysis with reasoning enabled to ensure the model has sufficient space to process visual data.
• Test the NVFP4 variant on ml.g7.24xlarge instances first to evaluate baseline performance before scaling to larger hardware configurations.