Amazon Nova

On this page 7

Amazon Nova is a family of foundation models developed by Amazon that covers text understanding, multimodal understanding, image generation, video generation and real-time speech, delivered primarily as managed services through Amazon Bedrock. In the landscape of AI models & assistants, it is the model line that Amazon builds itself, as distinct from the third-party models it also hosts, such as Claude, Llama and Mistral.

Amazon introduced the family at the end of 2024 and has extended it since with larger models, a speech-to-speech model, an agent toolkit and a second generation. Its positioning centers on price-performance, low latency and tight integration with Amazon Web Services rather than on consumer chat, which makes it primarily a building block for enterprise applications.

Models in the Amazon Nova family

Amazon Nova is organized into understanding models, which read inputs and produce text, and creative content generation models, which produce images or video. The first generation also gained a speech model and an agent toolkit, so the family spans most mainstream generative AI tasks under one brand.

The table below lists the first-generation members and their roles. Context lengths and capabilities change with updates, so the figures describe the launch generation.

ModelInput and outputMain role
Nova MicroText in, text out, 128K contextLowest latency and cost for text tasks
Nova LiteText, image and video in, text out, 300K contextFast, inexpensive multimodal processing
Nova ProText, image and video in, text out, 300K contextBalanced accuracy, speed and cost
Nova PremierMultimodal in, text out, about 1M contextMost capable model, teacher for distillation
Nova CanvasText and images in, image outImage generation and editing
Nova ReelText and images in, video outShort video clip generation
Nova SonicSpeech in, speech outReal-time voice conversation
First-generation Amazon Nova models

In late 2025 Amazon announced a second generation, Nova 2, with reasoning-capable Lite and Pro models, an updated Sonic model and an omni model that accepts and generates several modalities. Alongside it came Nova Forge, a service that lets organizations blend their own data into training checkpoints instead of fine-tuning only a finished model.

How Amazon Nova is accessed and customized

Amazon Nova is accessed mainly through Amazon Bedrock, the managed AWS service that exposes many foundation models behind a common API. Developers call Nova through the Bedrock Converse or InvokeModel interfaces, so switching between Nova and other hosted models often requires only a change of model identifier.

Bedrock also provides the surrounding features that matter in production: knowledge bases for retrieval-augmented generation, agents with tool use, guardrails for content filtering and sensitive-data handling, and batch inference for large offline jobs. Data sent to Bedrock is not used to train the base models, which is a key requirement for many regulated customers.

Customization options range from lightweight to deep, as the following table shows.

MethodWhat changesTypical use
Prompting and RAGOnly the input contextCompany knowledge, quick prototypes
Fine-tuningModel weights on labeled examplesDomain style, fixed output formats
Model distillationSmall model learns from a larger NovaCheaper model with near-teacher quality
Continued pretrainingWeights on large unlabeled corporaSpecialized vocabulary and domains
Nova ForgeData mixed into training checkpointsDeep specialization for large enterprises
Customization paths for Amazon Nova

Beyond the API, Amazon offers Nova Act, a toolkit for building agents that operate web browsers, and uses Nova models inside its own products, including parts of the Alexa+ assistant and Amazon Q.

How Amazon Nova compares with other model families

Amazon Nova competes mainly on cost per token, latency and AWS integration rather than on claiming the top position in every benchmark. Its understanding models are generally placed in the same tier as other providers’ fast and mid-sized models, while the most demanding reasoning workloads are often routed to frontier models available in the same Bedrock catalog.

FamilyProvider and distributionTypical differentiator
Amazon NovaAmazon, mainly via Amazon BedrockLow cost, AWS-native features
ClaudeAnthropic, own apps and major cloudsStrong reasoning, coding and long context
GeminiGoogle, own apps and Google CloudNative multimodality, Google integration
LlamaMeta, open weightsSelf-hosting and full weight control
MistralMistral AI, API and open weightsEfficient models, European provider
Amazon Nova compared with other model families

A practical consequence is that many teams use Nova as one tier in a model-routing strategy: Micro or Lite for classification and extraction, Pro or Premier for complex tasks, and a third-party frontier model where its quality justifies the higher price. Benchmark rankings shift with each release, so evaluation on the actual workload remains the reliable way to choose.

Typical use cases and model selection

Amazon Nova suits high-volume, latency-sensitive workloads where small differences in quality matter less than throughput and cost. Examples include summarizing support tickets, extracting fields from documents, captioning product images and generating marketing visuals.

The decision table below maps common requirements to a reasonable starting model.

RequirementStarting modelReason
Text classification at scaleNova MicroLowest latency and price
Invoice or receipt extractionNova LiteReads images and PDFs cheaply
Multistep analysis of reportsNova ProBetter accuracy on complex instructions
Long documents or videoNova PremierLargest context window
Product imageryNova CanvasGeneration, inpainting and background removal
Voice customer serviceNova SonicLow-latency speech-to-speech
Choosing a starting model

Limitations and responsible use

Amazon Nova has practical limitations that teams should weigh before committing. The models are closed-weight and tied to AWS, so they cannot be self-hosted elsewhere, and regional availability varies by model. Like all large models, Nova can produce confident but incorrect output, which requires grounding and human review in critical applications.

  • Vendor lock-in: workflows built around Bedrock features are harder to move to other platforms.
  • Benchmark claims: launch comparisons are published by the vendor and should be verified independently.
  • Generated media: Canvas and Reel outputs carry an invisible watermark and content credentials, which aid provenance but do not remove legal review.
  • Model updates: new versions can change behavior, so production prompts need regression testing.

In simple terms

Amazon Nova is Amazon’s own set of AI models. Think of it as a toolbox with differently sized tools: a small, fast one for simple text jobs, larger ones that can also look at pictures and videos, and special tools that draw images, make short videos or talk with people by voice.

Companies rent these tools through Amazon’s cloud instead of building their own AI. The main appeal is that they are fast, relatively inexpensive and easy to plug into systems that already run on Amazon Web Services.

A practical example

A mid-sized online retailer receives about 40,000 supplier invoices per month as scanned PDFs. Its finance team already runs its systems on AWS and wants to stop typing invoice data by hand.

The engineering team builds a pipeline in Amazon Bedrock. Nova Lite reads each invoice and returns supplier name, invoice number, dates, line items and totals as JSON. A validation step checks that line items add up to the total; when they do not, the invoice is escalated to Nova Pro for a second attempt and then to a person if it still fails.

After a month of tuning, most invoices pass without human touch, and the finance team only reviews the flagged exceptions. Because the cheaper model handles the bulk of the volume, processing cost stays low, and the retailer keeps all documents inside its existing AWS account, which satisfies its auditors.