At re:Invent 2025, AWS announced the new Nova 2 series of self-built models, which includes four different positioning versions: Lite, Pro, Sonic, and Omni. It also launched the Nova Forge Open Training service and the Nova Act service, which is designed to build reliable web browsing agents.
This update will enable enterprises not only to "use" AI, but also to deeply "customize" and "deploy" AI systems that align with their own business operations. It has already been adopted by companies such as Cisco, Siemens, Reddit, and Sony.
The Nova 2 family: From voice dialogue to all-around multimodal
AWS has launched four Nova 2 series models at once, covering a full range of needs from lightweight applications to complex inference:
• Nova 2 Lite:It features a high-performance, cost-effective inference model capable of handling text, images, and video. AWS states that its performance and cost-effectiveness outperform Anthropic's Claude Haiku 4.5 and OpenAI's GPT-5 Mini in multiple benchmark tests, making it suitable for customer service chatbots and document processing.
• Nova 2 Pro:This is Amazon's most powerful in-house inference model to date, excelling in programming, long-term planning, and complex problem-solving. It supports knowledge distillation and can be used as a "teacher" to train smaller models. Its performance is touted as comparable to Anthropic's Claude Sonnet 4.5 and Google's Gemini 2.5 Pro.
• Nova 2 Sonic:Specializing in speech-to-speech, it supports real-time, low-latency multilingual conversations and can seamlessly switch between speech and text.
• Nova 2 Omni:This industry-first technology enables multimodal performance on a single model, simultaneously processing text, images, video, and voice input, and generating text and images. It can analyze up to 75 words, hours of audio, or long video content at once, making it ideal for marketing teams to generate a complete set of promotional materials in one place.
Nova Forge: Pioneering "Open Training" to enable enterprises to create their own "Novellas"
AWS launched the Nova Forge service to address the needs of enterprises that want to deeply integrate proprietary knowledge into AI models.
This service pioneered the "open training" model, allowing businesses to access pre-trained, mid-trained, and post-trained checkpoints of Nova models. This means that businesses can mix their own data with Amazon's dataset at various stages of training to create optimized variant models known as "Novellas".
Reddit's Chief Technology Officer, Chris Slowe, stated that through Nova Forge, Reddit is using a more consistent system to improve content moderation, replacing the previous use of multiple fragmented models.
Nova Act: A powerful tool for automating browser operations, boosting Hertz testing efficiency by 5 times.
Finally, Nova Act is a service specifically designed for building AI agents for web browsing.
Based on a customized Nova 2 Lite model, trained through reinforcement learning, Nova Act can perform interface operations such as clicking and typing in the browser, automating tasks such as updating customer relationship management (CRM) data, testing website functions, or submitting insurance claims.
AWS claims that Nova Act achieved 90% reliability in early customer workflow testing. Car rental company Hertz used the tool to automate end-to-end testing of its rental platform, reducing a quality assurance (QA) process that previously took weeks to just hours, resulting in a 5x increase in software delivery speed.
