one year on
Amazon unveils Nova family of multimodal AI models at re:Invent
AWS enters the frontier model race with four text-generating models, an image generator and a video generator, with Jassy saying the models are among the fastest in their class and among the least expensive to run.
At its re:Invent conference today, Amazon Web Services announced Nova, a new family of multimodal generative AI models. The lineup includes four text-generating models — Micro, Lite, Pro, and Premier — along with an image-generation model, Nova Canvas, and a video-generation model, Nova Reel. Micro, Lite, Pro, Canvas and Reel are available immediately on AWS Bedrock; Premier, described as the most capable and intended as a teacher model for fine-tuning, is expected in early 2025.
CEO Andy Jassy said the models have made “a tremendous amount of progress over the last four to five months” and are among the fastest and least expensive in their class. Micro has a 128K-token context window, while Lite and Pro handle 300K tokens; context windows are planned to expand to over 2 million tokens next year. Canvas offers background removal, color and layout controls, while Reel generates up to six-second videos from text or reference images, with a version producing two-minute videos coming soon. Jassy also previewed a speech-to-speech model for Q1 2025 and an any-to-any model for mid-2025.
Amazon is indemnifying customers against copyright risks but remains vague about training data, citing competitive advantage. Jassy claims the Nova models are among the fastest in their class and among the least expensive to run.
The record
One year later — open only if you can handle spoilers
Nova models gained modest adoption among AWS customers, but Amazon's big AI bet remained Anthropic's Claude. The speech-to-speech model arrived on schedule, while the any-to-any model faced delays and never achieved the multimodal fluency Jassy promised. By mid-2026, Nova was seen as a competent but not category-leading offering.