Ddos 
AWS announced the debut of its new self-developed Nova 2 model family at re:Invent 2025, unveiling four distinct variants — Lite, Pro, Sonic, and Omni — alongside Nova Forge, an open-training service, and Nova Act, a platform purpose-built for creating reliable web-browsing AI agents.
This wave of updates enables enterprises not merely to use AI, but to deeply customize and deploy systems that reflect their own operational logic. The ecosystem is already being adopted by Cisco, Siemens, Reddit, Sony, and others.
AWS introduced all four Nova 2 models at once, covering the full spectrum from lightweight workloads to advanced reasoning:
- Nova 2 Lite — a cost-efficient reasoning model capable of handling text, images, and video. According to AWS, it surpasses Anthropic’s Claude Haiku 4.5 and OpenAI’s GPT-5 Mini in both performance and price efficiency, making it ideal for customer-support bots and document processing.
- Nova 2 Pro — Amazon’s most powerful in-house reasoning model to date, adept at code generation, long-horizon planning, and complex problem-solving. It supports knowledge distillation, allowing it to serve as a “teacher” for smaller models, and is positioned as a rival to Anthropic’s Claude Sonnet 4.5 and Google’s Gemini 2.5 Pro.
- Nova 2 Sonic — dedicated to speech-to-speech interaction, offering realtime, low-latency multilingual conversation with seamless transitions between spoken and written modes.
- Nova 2 Omni — the industry’s first single model capable of executing fully multimodal behavior: processing text, images, video, and audio inputs while generating both text and images. It can analyze up to 750,000 words at once, hours of audio, or extended video content, making it particularly suited to marketing teams seeking an end-to-end asset-generation engine.
To meet enterprise demand for integrating proprietary knowledge into AI systems, AWS introduced Nova Forge.
This platform pioneers an open-training paradigm, allowing enterprises to access Nova model checkpoints at the pre-trained, mid-trained, and post-trained stages. This enables organizations to blend internal datasets with Amazon’s own training corpora across multiple phases, producing optimized derivative models known as “Novellas.”
Reddit CTO Chris Slowe noted that Nova Forge allows the company to unify and strengthen its content-moderation systems, replacing an assortment of disparate models with a more coherent framework.
Finally, Nova Act emerges as a specialized service for building autonomous web-browsing AI agents.
Built atop a customized Nova 2 Lite model and trained using reinforcement learning, Nova Act can perform browser-level actions such as clicking, typing, and navigating interface elements. It automates tasks like CRM updates, website functionality testing, and insurance-claim submissions.
AWS reports that Nova Act achieved 90% reliability in early workflow trials. Hertz, for instance, used the tool to automate end-to-end testing of its rental platform, compressing quality-assurance cycles that once required weeks into mere hours and accelerating software delivery by a factor of five.
Donate