Amazon Nova Multimodal Embeddings
Embedding modelCard summarizedThe first unified embedding model that supports text, documents, images, video, and audio through a single model
Source: AWS AI Service Card: Amazon Nova Multimodal Embeddings (current as of 28 Oct 2025)
Where to call it
- Amazon Bedrock
amazon.nova-2-multimodal-embeddings-v1:0
Good for
- Agentic RAG over documents with text, images and infographics, video and audio
- Multimodal semantic search (e.g. digital asset management, e-commerce)
- Clustering media assets without metadata
Overview
Nova MME
The first unified embedding model that supports text, documents, images, video, and audio through a single model
- Dimensions
- 3072
- Max input
- 8K tokens text
- Modes
- text embeddings, image embeddings
- Released
- 28 Oct 2025
Amazon Nova Multimodal Embeddings is a proprietary Bedrock model that maps text, documents, images, video and audio into one embedding space for multimodal semantic search, agentic RAG and clustering. It offers four output sizes (3072 default, 1024, 384, 256), sync and async APIs with built-in segmentation of long media, and an embedding-purpose setting. AWS's service card describes its evaluation approach and fairness controls but publishes no benchmark numbers in text. service card; Nova user guide
amazon.nova-2-multimodal-embeddings-v1:0- Typical price
- $0.135 / 1M tok
- Dimensions
- 3072 (default), 1024, 384 or 256
- Max input
- 8K tokens text; 30 s video/audio per segment (async API segments longer files)
- Languages
- Text in up to 200 languages
- Weights
- Closed
- Released
- 28 Oct 2025
- text embeddings
- image embeddings
- document-image embeddings
- audio embeddings
- video embeddings (with or without audio)
- multimodal embeddings