Amazon Nova

Amazon Nova is Amazon's family of foundation models, accessed mainly through Amazon Bedrock, used to generate and analyze text, images, video and speech in AI applications.

The current AWS guide (Nova 2) lists three models:

• Nova 2 Lite: takes text, images, video and documents and answers in text, with extended reasoning and three intensity levels.
• Nova 2 Sonic: a speech-to-speech model for real-time voice applications.
• Nova Multimodal Embeddings: generates embeddings for text, documents, images, video and audio.

The Nova version 1 documentation, still listed in Bedrock, describes other models: Micro, Lite, Pro and Premier, which take text, image and video input and return text, plus Nova Canvas, which generates images, and Nova Reel, which generates video. AWS itself notes that new features and updates are published in the Nova 2 guide.

There is also Nova Act, an AWS service for building and managing AI agents that automate workflows in browser interfaces.

To try the models without entering Bedrock, AWS points to nova.amazon.com.

Amazon Nova is for development teams building generative AI applications on Amazon Bedrock that need text, image, video, speech or embedding models from the same family.

Use cases that AWS documentation describes:

• Real-time voice applications, such as spoken-conversation support, with Nova 2 Sonic.
• Analysis of documents, images and video through text questions, with Nova 2 Lite.
• Search and retrieval over mixed content, with the multimodal embeddings, which accept text, documents, images, video and audio.
• Marketing and advertising, which AWS cites as the main use of the creative models Nova Canvas (images) and Nova Reel (video).
• Browser task automation, with Nova Act.

What to have before you start:

• Access to Amazon Bedrock, which is the documented path. The API uses the InvokeModel or Converse operations, and the console has playgrounds for Nova 2 Lite and Nova 2 Sonic.
• For voice, a client able to use Nova 2 Sonic's bidirectional streaming API.
• For the creative models, English prompts, according to the version 1 documentation.

Amazon Nova's main features are multimodal understanding models, real-time speech-to-speech, multimodal embeddings, and image and video generation, all accessed through Amazon Bedrock.

• Nova 2 Lite: text, image, video and document input, text output. It has extended reasoning with intensity control at three levels.
• Nova 2 Sonic: speech and text as input and output, through a bidirectional streaming API. The documentation lists English (US, UK, India and Australia), French, Italian, German, Spanish and Portuguese, and sets an 8-minute limit per connection.
• Nova Multimodal Embeddings: text, documents, images, video and audio. Short content uses the synchronous InvokeModel call, and long video uses the asynchronous StartAsyncInvoke call.
• Nova Canvas: image generation from text and images, with inpainting, outpainting and style transfer from reference images. It supports fine-tuning in Bedrock with your own data.
• Nova Reel: video generation at 1280x720 and 24 frames per second, up to 2 minutes, according to the version 1 documentation.
• Nova Act: a service for building and managing agents that automate workflows in browser interfaces and can escalate to a human supervisor.

Canvas and Reel accept English prompts, with limits of 1,024 and 512 characters.

Amazon Nova is worth it for teams already building on Amazon Bedrock that want text, voice, embeddings and media generation in one family, with the same API operations. For teams that do not use AWS, the documented access path means going through it.

Pros:

• It covers different modalities in one family: understanding of text, images, video and documents, speech, multimodal embeddings, and image and video generation.
• Nova 2 Sonic does speech-to-speech in a single model with a bidirectional streaming API.
• Nova Canvas supports fine-tuning in Bedrock with your own data, to generate images aligned with a brand.
• The multimodal embeddings accept five content types in a single model.

Cons:

• The documented access path is Amazon Bedrock, which ties the application to AWS.
• Nova 2 Sonic limits each connection to 8 minutes.
• The creative models accept English prompts and have length limits: 1,024 characters on Canvas and 512 on Reel, according to the version 1 documentation.
• The documentation is split between version 1 and Nova 2, and AWS directs readers to the Nova 2 guide for new features.

The main alternatives to Amazon Nova as an AI model family are NVIDIA, IBM Granite, AI21 Labs and Ai2, and all of them offer models with published weights or private deployment.

• NVIDIA: the open Nemotron model family, with weights on Hugging Face, plus NeMo for data, fine-tuning, retrieval and guardrails and NIM microservices for deployment.
• IBM Granite: language, vision, speech, embedding and guardrail models with open weights under the Apache 2.0 license, according to IBM, and also available through watsonx.ai.
• AI21 Labs: Jamba models, a hybrid Transformer-Mamba architecture, with a 256K-token window and a VPC or on-premises deployment option, plus Maestro for agents with requirements validation.
• Ai2 (Allen Institute for AI): Olmo models published with training data, code and checkpoints, and Molmo vision models.

In practice the difference is access: Amazon Nova is documented for use through Amazon Bedrock, while the four alternatives above publish weights on Hugging Face, some with their own APIs.