The Amazon Nova Family of Models: Technical Report and Model Card

News Release Summary

Amazon has released a suite of new AI foundation models called Amazon Nova, spanning text, image, and video generation, and detailed their design and performance in an accompanying technical report. The lineup includes three text-and-multimodal understanding models — Nova Pro, Nova Lite, and Nova Micro — along with Nova Canvas for image generation and Nova Reel for video generation. The understanding models are built on the Transformer architecture and trained on multilingual data covering more than 200 languages, using a pipeline that moves from pretraining through supervised fine-tuning and reinforcement learning from human feedback via methods like DPO and PPO. On standard benchmarks, the models trade blows with comparable offerings from Anthropic, Google, and OpenAI: Nova Micro, the smallest text-only model, holds its own against similarly sized competitors on math and reasoning tasks, while the multimodal Pro and Lite models lead or place second on video captioning and several web-agent navigation tests. The image and video generation models, Canvas and Reel, use latent diffusion architectures and were evaluated through a mix of automated metrics and human preference studies. Notably, the report emphasizes practical tradeoffs — Nova Micro produces responses at 210 tokens per second compared to Claude 3.5 Sonnet's 57 — positioning the family as competitive on price-performance grounds rather than raw capability alone. The report also documents responsible AI measures including internal and external red-teaming and automated adversarial testing. The release matters because it gives developers and researchers a detailed public accounting of how a major cloud provider's proprietary model family compares to frontier competitors across a broad range of real-world tasks.