Amazon has announced Amazon Nova, a family of foundation models designed for generative AI tasks. The announcement, made during AWS re:Invent, highlights the models' capabilities in tasks such as document and video analysis, chart comprehension, video content generation, and AI agent development.
Amazon Nova includes two main categories: understanding models and content generation models.
The understanding models currently available are:
Amazon Nova Micro: Optimized for speed and cost, this model processes only text with low latency.
Amazon Nova Lite: A multimodal model capable of processing images, video, and text, generating text output at a lower cost.
Amazon Nova Pro: Designed for high-capacity tasks, this model can process up to 300,000 input tokens, allowing it to handle large codebases and complex data.
A fourth understanding model, Amazon Nova Premier, is in development and is expected to launch in early 2025. It will focus on complex reasoning tasks and serve as a teacher model for custom applications.
The content generation models include:
Amazon Nova Canvas: Generates images with editing features such as inpainting, outpainting, and background removal.
Amazon Nova Reel: Creates short videos based on text or image prompts, offering customization options for visual style and pacing.
The launch of Amazon Nova has received significant attention from the developer community, with many praising its balance of performance and cost.
Simon Willison, a well-known software engineer and co-creator of Django, shared his impressions on X, stating,
I spent some time yesterday exploring the new Amazon Nova LLM family, and I'm really impressed. They're price and quality competitive with Google Gemini – and Nova Micro is now the cheapest model from any of the major vendors (cheaper even than Gemini 1.5 Flash-8B).
A Reddit user, FarkCookies, commented:
Judging by placement, it is another homegrown model from AWS? Since it is next to Titan.
Amazon Nova models are built with a focus on safety and security, incorporating safeguards against misuse such as misinformation and harmful content. Generated outputs include digital watermarking to promote responsible use.
The models support over 200 languages, enabling global applications across diverse regions. For deployment, Amazon Bedrock offers features such as real-time streaming, batch processing, and integration with tools like Amazon Bedrock Knowledge Bases and Guardrails. Developers can explore these models through the Amazon Bedrock console, integrate them into workflows, and optimize their performance using detailed monitoring and feedback tools.
