Presenting the Amazon Nova foundation models: industry-leading pricing performance and frontier insight
Amazon Nova, a brand-new generation of cutting-edge foundation models (FMs) that are only accessible on Amazon Bedrock and offer industry-leading pricing performance and frontier intelligence.
Amazon Nova: What is it?
Available only on Amazon Bedrock, Amazon Nova is a new generation of state-of-the-art (SOTA) foundation models (FMs) that offer industry-leading price-performance and frontier intelligence.
For practically any generative AI task, Amazon Nova can reduce latency and expenses. With Amazon Nova, you can create sophisticated AI agents from a variety of intelligence classes tailored for enterprise workloads, analyze intricate documents and videos, comprehend charts and diagrams, and create captivating video content.
With two categories of models—understanding and creative content generation, it offers the intelligence and flexibility you require, whether you’re building AI assistants that can comprehend and act on visual information, developing document processing applications that must process text and images, or producing marketing content at scale.
To produce text output, Amazon Nova understanding models can take in text, images, or videos. Text and picture inputs can be used by Amazon creative content generation models to produce images or videos.
Model comprehension: Visual and textual intelligence
Three understanding models are included in the Amazon Nova models (a fourth will be released soon) and are intended to address various needs:
Amazon Nova Micro: It is an extremely inexpensive text-only model with the lowest latency answers among the Amazon Nova model’s models. Amazon Nova Micro is best at tasks like text summarization, translation, content classification, interactive discussion and brainstorming, basic mathematical reasoning, and coding. It has a context length of 128K tokens and is optimized for speed and cost. To improve accuracy, Amazon Nova Micro also allows customization on proprietary data through model distillation and fine-tuning.
Amazon Nova Lite: It is a very affordable multimodal model that processes text, video, and image inputs incredibly quickly to produce text output. Amazon Nova Lite is highly accurate at document analysis, visual question responses, and real-time customer interactions. The model can analyze up to 30 minutes of video or numerous photos in a single request, and it can interpret inputs up to 300K tokens in length. Using methods like model distillation, Amazon Nova Lite may be tuned to provide the optimum quality and price for your use case. It also enables text and multimodal fine-tuning.
Amazon Nova Pro: For a variety of jobs, Amazon Nova Pro is a very powerful multimodal model that offers the best accuracy, speed, and cost combination. Setting new benchmarks in multimodal intelligence and agentic workflows that need to use APIs and tools to finish complex processes, Amazon Nova Pro can process up to 300K input tokens. It attains cutting-edge results on important benchmarks like as video comprehension (VATEX) and visual question answering (TextVQA).
Amazon Nova Pro is very good at analyzing financial papers and shows great ability to process both textual and visual information. It can process code bases with more than 15,000 lines of code when given a 300K token input context. In order to distil unique variations of Amazon Nova Micro and Lite, Amazon Nova Pro also functions as a teaching model.
Amazon Nova Premier: It is the greatest teacher for distilling custom models and its most powerful multimodal model for challenging reasoning tasks. Training for Amazon Nova Premier is ongoing. It is to be available in early 2025.
Amazon Nova understanding models perform exceptionally well in agentic applications, function calling, and retrieval-augmented generation (RAG). This is demonstrated by the Amazon Nova model scores in VisualWebBench, Berkeley Function Calling Leaderboard (BFCL), Mind2Web, and the Comprehensive RAG Benchmark (CRAG) evaluation.
Amazon Nova’s customization features are what give it its special power for businesses. Consider it similar to making a suit: you begin with a fine foundation and modify it to meet your precise requirements. You may adjust the models using text, image, and video to better fit your brand voice, comprehend the jargon used in your sector, and optimize for your particular use cases. To better comprehend legal jargon and document formats, for example, a law firm may modify Amazon Nova.
Generating creative content: Making ideas a reality
Two models for producing creative material are also part of the Amazon Nova models:
Amazon Nova Canvas: A cutting-edge picture generating model, Amazon Nova Canvas offers comprehensive editing capabilities including inpainting, outpainting, and background removal, and produces studio-quality images with exact control over style and content. When it comes to human evaluations and important benchmarks like ImageReward and text-to-image fidelity evaluation with question answering (TIFA), Amazon Nova Canvas performs quite well.
Amazon Nova Reel: It is a cutting-edge model for creating videos. You can adjust visual style and pacing, create short movies using text prompts and graphics, and create high-quality video material for promotion, advertising, and entertainment with Amazon Nova Reel. When it comes to human assessments of video consistency and quality, Amazon Nova Reel performs better than current versions.
To encourage responsible AI use, all these models come with integrated safety features, and models for creating creative content have watermarking capabilities.
Throughout the model development process, Amazon Nova models are constructed with an emphasis on client safety, security, and trust, giving you peace of mind and a sufficient degree of control to support your particular use cases.
You have the controls you need to use AI properly with its extensive safety protections and content moderation tools. Digital watermarking is included in every created image and video.
The Amazon Nova foundation models are constructed with safeguards commensurate with their enhanced capabilities. In order to prevent the spread of false information, child sexual abuse material (CSAM), and chemical, biological, radiological, or nuclear (CBRN) hazards, Amazon Nova has expanded its safety protocols.
Things to be aware of
Amazon Bedrock in the US East (N. Virginia) AWS region offers Amazon Nova models. Through cross-region inference, Amazon Nova Micro, Lite, and Pro are also accessible in the US East (Ohio) and US West (Oregon) areas. Pay-as-you-go pricing is used for Amazon Bedrock, as is customary. Check out Amazon Bedrock price for additional details.
Your language is understood by the latest generation of Amazon Nova understanding models. English, German, Spanish, French, Italian, Japanese, Korean, Arabic, Simplified Chinese, Russian, Hindi, Portuguese, Dutch, Turkish, and Hebrew are among the more than 200 languages in which these models can comprehend and produce information. As a result, you can create apps that are genuinely global without having to worry about maintaining distinct models for various geographical areas or dealing with language issues. English prompts are supported by Amazon Nova models for creative content creation.
You’ll learn that Amazon Nova can manage ever more difficult jobs as you investigate it. These models can read long texts of up to 300K tokens, recognize up to 30 minutes of video content, analyze numerous photos in a single request, and produce images and videos at scale from natural language. This makes these models appropriate for a wide range of commercial use cases, including deep analysis of corporate paperwork, rapid customer service interactions, and the generation of assets for social media, e-commerce, and advertising.
Deployment and scaling are made simple via integration with Amazon Bedrock. You can use tools like Amazon Bedrock Guardrails to encourage appropriate AI use, Amazon Bedrock Agents to automate intricate workflows, and Amazon Bedrock Knowledge Bases to enrich your model with private data. Batch processing for heavy workloads, real-time streaming for interactive applications, and comprehensive performance monitoring are all supported by the platform.
