Amazon Unveils Its Nova AI Foundation Models at AWS Re:Invent 2024

At the AWS Re:Invent 2024 developer conference, Amazon introduced its own family of foundation models for generative AI, called "Amazon Nova." These models, available via AWS Bedrock, offer capabilities ranging from document analysis to creating complex AI agents, with lower latency and cost for generative AI tasks.

Nova Family Models

The Nova family includes specialized models for various tasks:

• Nova Micro: The simplest model, designed for text prompts and text-only responses.
• Nova Lite: Capable of handling images and videos up to 30 minutes long, but producing only text-based responses. Amazon's CEO, Andy Jassy, compared it to OpenAI’s GPT4o-Mini.
• Nova Pro: A high-performance multimodal model offering the best combination of accuracy, speed, and cost. It processes up to 300,000 tokens and supports text, video, image, financial documents, and code analysis, with outputs in text.

These models are currently available on AWS Bedrock but only in select US regions. They support over 200 languages, with particular strength in German, French, and Italian.

Looking Ahead: More Nova Versions in 2025

Amazon plans to launch "Nova Premier" in early 2025, designed for complex reasoning tasks and user-defined AI models, enabling pre-trained models to be retrained for specific applications.

Creative Content Generation Models

Additionally, Amazon introduced creative content generation models:

• Nova Canvas: Generates images from text prompts.
• Nova Reel: Produces short videos, initially up to 6 seconds, with plans to extend this to 2 minutes. AWS claims Nova Reel outperforms existing models in human evaluations of video quality and consistency.

These creative models are also available in select US AWS cloud regions and currently support only English prompts.

Upcoming Features for Nova

In Spring 2025, Amazon will release a version of Nova that can understand and generate spoken language. Later in 2025, a "any-to-any" version of Nova will be introduced, capable of processing and generating any type of input.

Source : ICTjournal