What's new in Amazon Nova 2
Amazon Nova 2 introduces significant enhancements across understanding, creative and speech capabilities. The following sections describe the key new features and improvements.
New models
-
A new multimodal model that combines understanding and generation capabilities. processes text, images, video and audio inputs and generates text and image outputs in a single unified model.
- Nova 2 Lite and
-
Enhanced models that offer advanced reasoning with extended thinking support, three-level intensity control and multimodal understanding optimized for production-scale agentic workflows.
- Nova 2 Sonic
-
An upgraded conversational speech model with improved speech understanding, natural language processing and voice generation capabilities across seven languages.
- Nova Multimodal Embeddings
-
A multimodal embedding model that processes text, image, document, video and audio inputs and generates embeddings.
New features and capabilities
Topics
Nova Multimodal Embeddings
Nova Multimodal Embeddings supports text, documents, images, video and audio through a single model, enabling cross-modal retrieval applications. Nova Multimodal Embeddings maps each of these content types into a unified semantic space, enabling you to conduct unimodal, cross-modal and multimodal vector operations, powering applications such as agentic retrieval-augmented generation (RAG) and multimodal semantic search. retrieval-augmented generation (RAG) and multimodal semantic search.
Extended thinking and reasoning
Nova 2 Lite, support extended thinking, which allows the models to spend more time reasoning through complex problems before generating responses. This capability improves accuracy for multi-step reasoning tasks, including agentic workflows with multiple tools, advanced mathematics, complex planning and code generation.
Built-in tools
Amazon Nova 2 includes built-in tools that extend model capabilities without requiring external integrations:
-
Web grounding – Accesses real-time information from the web to provide up-to-date responses and reduce hallucinations.
-
Code interpreter – Executes Python code to perform calculations.
AI agent building
Amazon Nova 2 models are optimized for building AI agents. The models provide improved tool use, better reasoning for multi-step tasks and enhanced ability to maintain context across complex agent workflows.
Improved document understanding
Nova 2 Lite and provide enhanced document processing capabilities with better understanding of complex document layouts, tables, charts and multi-page documents. The models extract information more accurately from PDFs, spreadsheets and other document formats.
Enhanced video understanding
Nova 2 Lite and offer improved video analysis capabilities, including better visual perception, temporal understanding, action recognition and the ability to process longer video sequences with higher accuracy.
Model customization
Nova 2 Lite supports supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) on Amazon Bedrock and SageMaker AI AI, allowing you to adapt Amazon Nova 2 to your specific business needs.
Amazon Nova Forge
Amazon Nova Forge is a first-of-its-kind service that offers organizations the easiest and most cost-effective way to build their own frontier models using Amazon Nova.
Next steps
-
To learn about Amazon Nova models and capabilities, see What is Amazon Nova 2?.
-
To start using Amazon Nova 2.0, see Getting started with Amazon Nova 2.
-
To explore core inference features, see Core inference.
