Nova 2 Documentation
What Amazon Nova 2 is, which models are available, and how to start building multimodal applications with up to 1 million tokens of context.
What is Nova 2?
Amazon Nova 2 is a family of multimodal foundation models that process text, images, video, documents and speech in a single request. Nova 2 models support up to 1 million tokens of context and can generate up to 65,536 tokens in a single response.
With Nova 2 you can build interactive chatbots, analyze documents and videos, create agents with extended reasoning, and develop voice-enabled applications.
Models
| Model | Input modalities | Output modalities | Best for |
|---|---|---|---|
| Nova 2 Lite | Text, images, video, documents | Text | High-volume applications that prioritize speed and cost efficiency |
| Nova 2 Sonic | Speech, text | Speech, text | Voice-enabled applications with fast response times |
| Nova Multimodal Embeddings | Text, images, documents, video, audio | Embeddings | Semantic search, recommendation systems and similarity matching |
Context window and output
- Context window: up to 1,000,000 tokens
- Maximum output: 65,536 tokens per response
- Reasoning: models with reasoning capabilities perform extended thinking and show step-by-step analysis before answering
How Nova 2 works
- Your application sends a request with your input and configuration parameters.
- The model processes your input, applying reasoning when it is configured.
- The model generates a response and returns it to your application.
- You can enhance responses with RAG, built-in tools such as web grounding and code interpreter, or custom fine-tuning.
Key concepts
Foundation models — pre-trained AI systems available in different sizes and capabilities that you access through an API.
Inference — the process of sending a request to a model and receiving a generated response.
Reasoning — extended thinking that lets models break down complex problems and show their step-by-step analysis.
Multimodal — the ability to process and understand multiple input types together in a single request.
RAG (Retrieval-Augmented Generation) — a technique that combines model responses with your own data sources for more accurate, contextual answers.
Start creating
Sign in to the Nova 2 chat experience, or compare credit packs and plans on the pricing page.