Amazon Launches New Nova Models, Pioneers 'Open Training' with Nova Forge

TMTPOST -- Amazon.com Inc. unveiled a major expansion of its Nova AI portfolio at AWS re:Invent, introducing four new models that deliver competitive price-performance across multiple domains and pioneering an "open training" approach through Nova Forge that allows organizations to build customized frontier models using proprietary data. The announcements represent Amazon's latest effort to distinguish itself in the crowded AI market, where it has struggled to match its cloud computing dominance.

Credit：Amazon

The new Nova 2 family includes models spanning everyday reasoning tasks to complex multimodal processing. Amazon also launched Nova Act, a service for building browser-based AI agents that achieves 90% reliability on customer workflows. Tens of thousands of companies are already using Nova models across diverse applications, according to the company.

Rohit Prasad, who leads Amazon's model development and Artificial General Intelligence team, pushed back against industry benchmark standards. "The real benchmark is the real world," he said in an interview, expressing confidence the new models would prove competitive. Companies including Cisco, Siemens, Reddit, and Hertz are deploying the technology for applications ranging from content moderation to automated testing.

The expansion aims to attract customers seeking cost-effective alternatives to market leaders while offering new customization capabilities that go beyond surface-level integration of proprietary knowledge.

Nova 2 Models Balance Speed, Cost and Intelligence

The Nova 2 family comprises four distinct models optimized for different use cases. Nova 2 Lite serves as a fast, cost-effective reasoning model for everyday workloads, processing text, images, and videos to generate text. It delivers equal or better performance on 13 out of 15 benchmarks compared to Claude Haiku 4.5, equal or better on 11 out of 17 benchmarks versus GPT-5 Mini, and equal or better on 14 out of 18 benchmarks against Gemini Flash 2.5. The model excels at document processing, video information extraction, code generation, and multi-step workflow automation.

Nova 2 Pro represents Amazon's most intelligent reasoning model, handling text, images, video, and speech inputs to generate text. Designed for complex tasks like agentic coding, long-range planning, and sophisticated problem-solving, it performs equal or better on 10 out of 16 benchmarks compared to Claude Sonnet 4.5 and equal or better on 15 out of 19 benchmarks versus Gemini 2.5 Pro. Both Lite and Pro models feature built-in web grounding and code execution capabilities, enabling them to search current information and run code directly rather than relying solely on training data.

Nova 2 Sonic functions as a speech-to-speech model for real-time conversational AI, featuring expanded multilingual support, expressive voices, and a one-million token context window. It handles tasks asynchronously, allowing users to continue natural conversations while actions complete in the background, and integrates with Amazon Connect and telephony providers like Vonage and Twilio.

Nova 2 Omni breaks new ground as an industry-first unified multimodal model that processes text, images, video, and speech inputs while generating both text and images. It can handle up to 750,000 words, hours of audio, long videos, and hundred-page documents simultaneously, eliminating the cost and complexity of connecting multiple specialized models.

Nova Forge Enables Custom Model Building

Nova Forge addresses a longstanding compromise facing organizations embedding AI: superficial customization of proprietary models, training open-weights models without original training data risking capability regression, or building from scratch at enormous expense. The service provides exclusive access to pre-trained, mid-trained, and post-trained Nova model checkpoints, allowing customers to blend proprietary data with Amazon Nova-curated datasets at every training stage.

Organizations can build their own optimized Nova variants—called "Novellas"—that combine Nova's foundational knowledge and reasoning with deep understanding of specific business contexts. Customers can start with Nova 2 Lite immediately, while Nova Forge users gain early access to Nova 2 Pro and Nova 2 Omni.

The service offers additional capabilities including reinforcement learning "gyms"—synthetic environments where models learn from simulated scenarios—and synthetic data-based distillation for creating smaller, faster models that maintain intelligence at lower cost. Reddit is using Nova Forge to improve content moderation with a unified system. "We're replacing a number of different models with a single, more accurate solution that makes moderation more efficient," said Chris Slowe, Reddit's chief technology officer.

Companies including Booking.com, Nimbus Therapeutics, and Sony are building custom models with Nova Forge for their unique requirements.

Nova Act Automates Browser-Based Workflows

Nova Act provides a new service for building and deploying AI agents that automate browser-based tasks, powered by a custom Nova 2 Lite model. The service delivers 90% reliability on early customer workflows through reinforcement learning training on thousands of tasks across hundreds of simulated web environments, excelling at UI-based workflows like CRM data updates, website functionality testing, and health insurance claims submission.

Developers can prototype agents in minutes using a no-code playground with natural language prompts, refine them in IDEs like VS Code, and deploy to AWS with comprehensive management tools. Hertz accelerated software delivery by 5x using Nova Act to automate end-to-end testing across its rental platform, transforming weeks of work into hours. Startup Sola Systems integrated Nova Act to automate hundreds of thousands of monthly workflows including payment reconciliation and shipment coordination.

The Nova Act console provides production-scale management and monitoring for deployed agent fleets.