Vibeleaderboard
Amazon

ai lab

Amazon

Amazon matters because it can combine original models, custom accelerators, a large cloud platform, enterprise controls, and internal deployment surfaces. Nova Forge also tests a deeper form of enterprise model customization than post-training alone. The useful judgment is not whether Amazon owns every layer, but whether integration across those layers produces models and agent systems that customers choose on evidence rather than procurement convenience.2,6,7,8,10

Official site ↗Models in the Atlas →

Profile

Overview

An original model program inside Amazon

Amazon's Nova foundation models are built by the Amazon AGI organization and partner teams inside the larger company. Amazon introduced the first Nova family through AWS in December 2024, with text, multimodal understanding, image, and video models. The program sits alongside older Amazon research in Alexa, recommendation, robotics, and machine learning, but Nova is the canonical original foundation-model line represented by this dossier.2,1

Nova, Bedrock, and AWS are different layers

Amazon delivers Nova through Amazon Bedrock, AWS's managed service for models from Amazon and outside developers. Bedrock is therefore a distribution and governance layer, not another name for the Nova lab. The wider stack includes Trainium and Inferentia accelerators, data-center infrastructure, and enterprise identity and storage controls. Amazon can optimize across these layers, while customers can still choose models from Anthropic, Google, OpenAI, Meta, Mistral, Cohere, and other labs through Bedrock.10,3,8

Agents and training-stage customization

The model portfolio expanded with Nova Act, a browser-action system, and the Nova 2 family. Nova 2 added text reasoning, speech, and an Omni preview designed to understand text, images, video, and speech while producing text and images. Nova Forge lets selected organizations combine proprietary data with Amazon-curated data at several training stages and access checkpoints rather than limiting customization to ordinary fine-tuning after a finished base model.4,6,7

A full-stack program under new leadership

Leadership changed after that expansion. Rohit Prasad led the creation of Nova and the Amazon AGI organization, then announced his departure at the end of 2025. Peter DeSantis took responsibility for Amazon's AGI models alongside custom silicon and quantum computing, while Pieter Abbeel took charge of the frontier team that builds the base model. Amazon reported strong demand for Trainium3 and wider Bedrock model coverage in 2026, but public evidence still describes infrastructure adoption more clearly than independent Nova model preference or sustained use.9,10

Notable contributions

  1. 01Training-stage enterprise customizationNova Forge gives customers access to model checkpoints and allows proprietary data to be mixed into several stages of training. Amazon describes this as open training. The notable contribution is the managed customization depth, not a claim that Amazon invented continued pretraining.6,7
  2. 02A production-oriented browser-action systemNova Act pairs a specialized browser-action model with an SDK and AWS service controls intended to move workflows from prototypes into scheduled or headless production tasks. Its benchmark claims remain primarily Amazon-reported.4,5
  3. 03An integrated model, silicon, and cloud stackAmazon connects Nova models to Bedrock and its Trainium infrastructure while preserving access to third-party models. This gives it control over more of the serving economics and deployment path than a standalone lab, although infrastructure integration does not by itself establish superior model capability.8,10