Amazon introduces the Nova foundation models

Background

Amazon's own models had previously been offered in Amazon Bedrock under the Titan name. On 3 December 2024 Amazon introduced a new generation of its foundation models, called Nova, at AWS re:Invent.^[1][2]

What happened

On 3 December 2024, at AWS re:Invent, Amazon introduced Amazon Nova, a new generation of foundation models.^[1][2] The launch included Nova Micro, Nova Lite and Nova Pro for understanding tasks, and Nova Canvas and Nova Reel for image and video generation.^[1]

The most capable model, Nova Premier, was announced for early 2025.^[1][2] At launch it was described as a future addition rather than an available model.

Key details

The models were priced well below comparable options. Amazon said they were at least 75% cheaper than the best-performing models in their class on Bedrock.^[2]

The models are available exclusively in Amazon Bedrock, and they support fine-tuning and distillation. Distillation is the process of training a smaller model to imitate a larger one, so a company can produce a cheaper model that keeps part of a bigger model's behaviour.^[1][2]

Context windows were set at 128K tokens for Micro and 300K tokens for Lite and Pro. The context window is the amount of text a model can take into account in one request.^[1]

Impact

The launch gave Amazon a new generation of models that covers both understanding and generation, in text, images and video, all sold through one managed service. It also set the terms on which the models would be sold: Bedrock only, with customers able to adapt the models to their own data.^[1][2]

For readers, the event is a clear example of how a major cloud provider builds its own model range. It starts with a few models in different sizes and types, and adds a more capable model later.^[1][2]

The launch also used terms that may be unfamiliar. A foundation model is a large model trained on broad data that can then be adapted to many tasks. Fine-tuning means further training a model on a customer's own examples. Distillation means training a smaller model to imitate a larger one. A context window is measured in tokens, and a longer window lets a model consider more text at once.^[2]