NVIDIA launches 120-billion-parameter multimodal model on Amazon SageMaker JumpStart.

NVIDIA launches 120-billion-parameter multimodal model on Amazon SageMaker JumpStart. The model supports image, video and text input.

The Nemotron 3.5 Super VL supports image, video and text input. It includes an MTP head and is used in document intelligence, GUI understanding, video question answering and multimodal agents. It can be deployed on Amazon EKS and SageMaker HyperPod EKS. The model is part of NVIDIA's AI ecosystem, which also includes Nemotron 3.5 Transcribe, a speech recognition model with 30 billion parameters. Nemotron 3.5 Transcribe combines transcription, translation and speaker recognition in up to 40 languages. The multimodal language model is an evolution of the Nemotron series, which already includes models with 100 billion parameters.

The architecture of Nemotron 3.5 Super VL is hybrid, combining Mamba-2, MoE and attention. It also uses a RADIO vision encoder and BF16 and NVFP4 checkpoint formats. NVIDIA unveiled the model in partnership with Amazon SageMaker JumpStart, a platform that provides access to pre-trained models. Availability in Brazil has not yet been confirmed, but the model can be used by developers who need text, image, document and video understanding in complex applications.

NVIDIA Nemotron 3.5 Super VL is a multimodal language model with 120 billion parameters, capable of processing image, video and text input. The number of parameters affects the model's ability to understand complex contexts and perform tasks such as document understanding, graphical interface analysis and answering questions about videos. On the other hand, the 12 billion active parameters per token determine the model's efficiency during execution.

This means that, even with a high total number of parameters, the model can maintain optimized real-time performance. The combination of reasoning and tool use allows the model not only to understand, but also to carry out actions based on multimodal data. Availability in Brazil has not yet been confirmed, but the model can be deployed on Amazon EKS and SageMaker HyperPod EKS, making it easier to integrate into development environments.

The Super VL, in turn, expands this capability with 120 billion parameters, activating 12 billion per token. Its hybrid architecture, which combines Mamba-2, MoE and attention, makes it possible to process text, image, document and video in an integrated way. This makes it a versatile tool for tasks such as document intelligence and answering questions about videos.

The version of the model on Amazon SageMaker JumpStart can also be deployed on Amazon EKS, expanding its scalability. NVIDIA already has an ecosystem of cloud AI models, highlighting its codesign strategy to optimize memory usage.

Specifications

Multimodal architecture expands contextual understanding

The NVIDIA Nemotron 3.5 Super VL replaces the traditional text-processing approach with a multimodal architecture that handles image, video and text simultaneously. The inclusion of the MTP head allows the model to process multiple input modalities in an integrated way, facilitating tasks such as GUI understanding and video question answering. Developers can build more versatile solutions without having to integrate multiple models.

Availability on SageMaker JumpStart simplifies access to the model, allowing technical teams to deploy advanced AI solutions without their own infrastructure. The lack of availability in Brazil is still an obstacle, but the model's versatility may attract users in regions with access to AWS services.

With information from the-decoder.com, huggingface.co and technews.tw.

Text written with AI from the sources listed and checked by the newsroom's automated rules. How we work

• #hardware
• #software