In the Bedrock console, the Nova Canvas, Reel, and Premier models now carry the “Heritage” label.
Behind the scenes, there appears to be a strategic shift. Amazon would be focusing on a frontier model that it could unveil at the end of 2026 during its re:Invent conference.
The initiative, apparently dubbed Frontier Model Research internally, would be led by Pieter Abbeel. He has served, among other roles, as co‑director of the AI Research Lab at UC Berkeley. In late 2024, Amazon acquired his robotics company Covariant.
To accompany this pivot, the Amazon AGI division was reorganized. It had been created in 2023 under Rohit Prasad, a veteran of the Alexa team (director of machine learning, among other duties). In late 2025, Amazon merged it with the Custom Dawn Liphardt and Quantum Computing divisions. Peter DeSantis oversees the entire operation. Having been with the company for about twenty years, he previously held roles including general manager of EC2 and SVP of infrastructure & support at AWS.
As part of this reorganization, Amazon closed its AGI lab in San Francisco. It had set it up in late 2024 after acquiring Adept AI Labs, a startup led by a former OpenAI and Google Brain veteran.
A First Generation by the End of 2024…
The Amazon Nova models officially debuted at re:Invent 2024. There were five at the time:
- Micro (text input and output; 128k context including 10k for output)
- Lite (text, image, and video input; text output; 300k context and 10k)
- Pro (same basic features as Nova Lite)
- Canvas (text and image input; image output)
- Reel (text and image input; video output)
In April 2025, Amazon added to the Nova lineup Nova Premier (text, image, and video input; text output; 1M context including 10k in output). And Nova Sonic (audio input, text and audio output; 300k).
… and a Second Wave by Late 2025
In October 2025, the family expanded with a vectorization model (Nova Multimodal Embeddings). The Nova 2 models arrived at re:Invent 2025:
- Lite (same modalities as the first generation, but with 1M context and an “extended thinking” mechanism)
- Pro (added audio input and moved to 1M context)
- Sonic (added text output and moved to 1M context)
- Omni (text, images, video, and audio input; text and image output)
From generation to generation, Amazon promised better alignment with agentic use (tool use, reasoning, context management…). It also added web anchoring and a native code interpreter.
Nova Omni was—and remains—an early access feature for Nova Forge customers. This service, offered on an annual subscription, allows continuing the training of Nova models from checkpoints by providing one’s own datasets.
What Remains in the Nova Catalog
The re:Invent 2025 also marked the full availability of Nova Act. This service enables the development of fleets of agents to automate browser-based UI workflows.
Apart from the models labeled “Heritage,” the Nova family currently includes:
- Micro
- Lite (generations 1 and 2)
- Pro (generations 1 and 2, the latter in preview)
- Sonic Generation 2 (the first one is labeled “Heritage”)
- Multimodal Embeddings
They are available—except for the vectorization model—on the Nova Playground.
Bedrock and SageMaker AI offer various customization options for these models (supervised fine-tuning, reinforcement learning, distillation…).
Amazon also has, in experimental form, a model dedicated to deeper research: Nova Deep Research.
