AI and models, cloud platforms, chips and computing hardwareConfirmedNVIDIA's powerful Nemotron 3.5 Super VL, a 120B-parameter multimodal vision-language model, is now available on Amazon SageMaker JumpStart. This release enables AWS customers to deploy advanced AI solutions for document intelligence, GUI understanding, video Q&A, and agentic applications, leveraging both image, video, and text inputs.
NVIDIA's Nemotron 3.5 Super VL model is now available on Amazon SageMaker JumpStart, expanding the portfolio of foundation models accessible to AWS customers. This model is a vision-language model from the Nemotron 3.5 family, designed to accept images, video, and text as input and generate text output.
The model is specifically engineered for enterprise applications such as document intelligence, graphical user interface (GUI) understanding, video question answering, and multimodal agentic applications. Its architecture features a 120-billion-parameter hybrid Mixture-of-Experts (MoE) language backbone, with 12 billion active parameters, paired with a vision encoder. It supports extensive context lengths, up to 1 million tokens.
Key capabilities include a built-in Multi-Token Prediction (MTP) head, which facilitates speculative decoding to accelerate text generation. The model's reasoning capabilities are configurable per request. It supports eight languages: English, French, Spanish, German, Japanese, Hindi, Korean, and Chinese. Nemotron 3.5 Super VL has undergone post-training with multimodal supervised fine-tuning across diverse data types, including image, video, document, chart, and GUI data, followed by GRPO reinforcement learning. This training enables it to deliver structured JSON output, support tool calling, and offer configurable reasoning for enterprise agent workflows.
Amazon SageMaker JumpStart provides two variants of the model. The first is an NVFP4 optimized version for inference, designed to run efficiently on G7 instances, which are equipped with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs. The second variant consists of BF16 reference weights, available on ml.p5en.48xlarge instances, intended for customization tasks such as supervised fine-tuning (SFT), reinforcement learning (RL), distillation, and domain adaptation. Customers can deploy these models with a few clicks through the SageMaker console or by using the SageMaker Python SDK.
The integration of NVIDIA's Nemotron 3.5 Super VL into Amazon SageMaker JumpStart represents a significant advancement for developers building sophisticated AI applications on AWS. The model's multimodal input capabilities—processing images, video, and text to generate text—address a critical need for enterprise solutions that interact with diverse data types. (Inference) This moves beyond traditional text-only large language models, enabling a broader spectrum of use cases where visual context is paramount, such as analyzing complex reports with embedded charts or understanding user interactions within a software interface. This capability is crucial for automating tasks that previously required human interpretation of mixed media.
The model's hybrid Mixture-of-Experts (MoE) architecture, featuring a 120-billion-parameter backbone with only 12 billion active parameters, is a strategic design choice for balancing model capacity with inference efficiency. (Inference) This design allows the model to leverage the expressive power of a very large parameter count while potentially reducing computational costs and latency during inference, as only a subset of experts are activated for any given input. For software engineers, this translates to more performant and cost-effective deployment of a high-capacity model in production environments, making advanced AI more accessible for real-world applications.
A standout feature is the support for context lengths up to 1 million tokens. (Inference) This exceptionally long context window is transformative for applications requiring deep understanding of extensive documents, prolonged conversations, or entire video transcripts. For example, in document intelligence, it allows the model to process and reason over entire legal contracts, financial reports, or technical manuals without fragmentation, maintaining coherence and accuracy across vast amounts of information. Similarly, for video question answering, it enables comprehensive analysis of extended video content, capturing nuances that shorter context windows would miss, which is vital for detailed content analysis.
The inclusion of a Multi-Token Prediction (MTP) head for speculative decoding directly addresses a common bottleneck in large language model deployment: generation speed. (Inference) Speculative decoding is a technique where a smaller, faster draft model proposes several tokens, which the larger model then efficiently verifies. By integrating this capability, Nemotron 3.5 Super VL aims to accelerate text generation, making it more suitable for real-time interactive applications where low latency is crucial for user experience. This is a practical engineering consideration for building responsive AI-powered interfaces or agents that need to provide quick responses.
The model's ability to deliver structured JSON output and support tool calling are foundational for building robust and reliable AI agents and integrating AI into existing software workflows. (Inference) Structured JSON output simplifies the parsing and programmatic consumption of model responses, eliminating the need for complex natural language processing on the output side. Tool calling empowers the model to interact with external APIs, databases, or other software components, extending its capabilities beyond text generation to perform actions in the real world. This is essential for creating autonomous agents that can execute tasks, retrieve information, or control systems based on user prompts, thereby enabling more complex automation.
The provision of two distinct variants on SageMaker JumpStart caters to different stages of the machine learning lifecycle. The NVFP4 optimized version, running on G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, is tailored for efficient, high-performance inference in production. (Inference) This optimization is critical for deploying large models at scale, where throughput and cost-efficiency are paramount. Conversely, the BF16 reference weights on ml.p5en.48xlarge instances offer flexibility for customization. (Inference) This allows enterprises to fine-tune, distill, or adapt the model with their proprietary data, ensuring the model's performance and behavior are precisely aligned with specific domain requirements or internal datasets, which is a common need for enterprise AI adoption and differentiation.
The comprehensive post-training regimen, involving multimodal supervised fine-tuning on diverse data (image, video, document, chart, GUI) followed by GRPO reinforcement learning, indicates a deliberate effort to enhance the model's real-world applicability and alignment. (Inference) This multi-stage training process aims to improve the model's understanding across modalities and refine its behavior to be more helpful, harmless, and honest, which are critical considerations for deploying AI in sensitive enterprise contexts. The configurable reasoning per request further allows developers to fine-tune the model's decision-making process for specific tasks, offering granular control over its operational logic and output.
The availability of Nemotron 3.5 Super VL on SageMaker JumpStart significantly lowers the barrier to entry for AWS customers. (Inference) By offering a 'few clicks' deployment experience, AWS abstracts away the complexities of infrastructure provisioning, model hosting, and scaling, allowing developers to focus on application logic rather than MLOps. This accessibility democratizes the use of advanced multimodal AI, enabling a wider range of organizations, including those with limited deep learning expertise, to leverage state-of-the-art foundation models for their business challenges and accelerate their AI initiatives.