AWS launched the second generation of its self-developed large model family Nova2 at re:Invent 2025, featuring four new products: Lite, a lightweight multimodal model; Pro, a complex reasoning agent; Sonic, a speech dialogue model; and Omni, with a context length of 750,000 words. The company also announced an interconnection agreement with Google Cloud, making it easier for customers to call competing models like Gemini across platforms.
The Nova2 series focuses on "industry-leading cost-effectiveness." Official pricing shows that Lite costs $0.045 per million tokens for input and $0.09 for output, approximately 50% of the level of similar models. AWS CEO Matt Garman said that the addition of multimodal capabilities, code generation, and agent task optimization aims to win "on cost rather than parameters."
What really caught developers' attention was the simultaneous release of the "Nova Forge" service: for an annual fee of $100,000, customers can participate in the pre-training, mid-training, or post-training stages, injecting private data to build cutting-edge custom models. Garman explained that early fine-tuning can weaken the model's reasoning ability, "Forge allows customers to customize models 'from scratch,' just like learning a language."
AWS revealed that internal e-commerce and Alexa teams have already used Forge to build business-scale models; external customer Reddit has used the service to train content moderation models, with the official claiming that it outperforms commercially available large models in internal evaluations. If customers need full assistance from Amazon engineers, additional consulting fees apply.
