Amazon introduces Nova 2 family with four models and custom variant creation service

Amazon has introduced an expansion of its Nova portfolio with four new artificial intelligence models, a service that allows organizations to create custom variants, and a tool for developing web automation agents. Nova 2 models are designed to balance speed, cost, and intelligence across multiple AI domains.

Nova 2 Lite is a fast and cost-effective reasoning model that processes text, images, and videos to generate text. It allows adjusting the amount of step-by-step reasoning performed before responding, balancing analysis depth with speed and cost.

Nova 2 Pro is Amazon's most advanced reasoning model, capable of processing text, images, video, and speech to generate text. It is designed for complex tasks such as agentic coding, long-range planning, and sophisticated problem-solving where maximum accuracy is required. The model can function as a "teacher" for knowledge distillation, transferring its capabilities to smaller, more efficient models.

Both models include integrated web search and code execution capabilities. Amazon notes that Nova 2 Lite and Pro match or exceed comparable models from Claude, OpenAI, and Google in most evaluated benchmarks.

Nova 2 Sonic is the speech-to-speech model that unifies text and speech understanding and generation for real-time conversational AI. It includes expanded multilingual support, a one-million token context window for sustained interactions, and handles tasks asynchronously allowing users to continue conversations while actions complete in the background.

Nova 2 Omni is a unified model that processes and generates text, images, video, and speech. It can handle up to 750,000 words, hours of audio, long videos, and hundreds of pages of documents simultaneously, eliminating the need to connect multiple specialized models.

Nova Forge allows organizations to create their own optimized Nova variants, called "Novellas," by combining their proprietary data with Nova's capabilities. The service offers exclusive access to model checkpoints at different training stages, allowing mixing of proprietary data with Amazon Nova-curated datasets. It also offers the ability to train using custom environments through reinforcement learning, create smaller models through synthetic data-based distillation, and access to responsible AI tools.

Nova Act is available as a service on AWS for building and deploying AI agents that automate tasks in web browsers. Powered by a custom Nova 2 Lite model and trained through reinforcement learning in simulated web environments, it achieves 90% reliability in workflows. Developers can prototype agents without code using natural language, refine them in IDEs like VS Code, and deploy to AWS.

Key points

• Amazon introduces Nova 2 family with four new models (Lite, Pro, Sonic, and Omni) alongside Nova Forge for creating custom variants and Nova Act for web automation
• Nova 2 Lite is a fast and cost-effective reasoning model that allows adjusting analysis depth to balance intelligence with speed and cost
• Nova 2 Pro designed for complex tasks and can function as "teacher" for knowledge distillation into smaller models
• Nova 2 Sonic unifies text and speech understanding and generation in real time with one-million token context window and asynchronous task management
• Nova 2 Omni is a unified model that processes and generates text, images, video, and speech, handling up to 750,000 words simultaneously
• Nova Forge enables creating custom model variants by integrating proprietary data at each training stage through "open training"
• Nova Act is a service for building AI agents that automate browser tasks, achieving 90% reliability