At the AWS re:Invent 2025 global conference, Amazon unveiled its groundbreaking Nova Forge service for building enterprise-specific AI models, alongside Nova Act for Agent development. Nova Forge introduces an "open training" approach, enabling businesses to integrate proprietary data early in the training phase to create highly customized Nova-optimized variants tailored to their needs.
Enterprises often face three suboptimal choices when infusing domain expertise into AI applications: limited fine-tuning of closed-source models, which only superficially incorporates business knowledge; continual training of open-source weight models without original training data, risking degradation in core capabilities like instruction-following; or building models from scratch at prohibitive cost and time. The ideal solution must combine cutting-edge model capabilities with deep integration of proprietary knowledge.
Nova Forge bridges this gap by merging enterprise data with Nova's advanced capabilities to create bespoke optimized variants called "Novellas." It provides access to Nova's pre-training, mid-training, and post-training model checkpoints, allowing customers to blend their data with Amazon Nova's curated datasets at each stage. This produces customized models that retain Nova's full knowledge and reasoning abilities while deeply understanding business contexts. Clients can now start building Novellas using Nova 2 Lite as the foundation.
Early adopters include Booking.com, Cosine AI, Nimbus Therapeutics, Nomura Research Institute, OpenBabylon, Reddit, and Sony, which are already leveraging Nova Forge to develop models aligned with their specific requirements.
Comments