Amazon Nova AI Explained: Models, Bedrock and What Each One Does

Amazon Nova AI, in plain English

A foundation model is a pre-trained AI system that an application can send a request to through an API. Amazon Nova is Amazon’s branded family of those systems. Depending on the model, the request might contain written text, a document, an image, a video clip, or spoken audio. The response might be text, speech, or a numerical representation used for search. That is why “Amazon Nova AI” is a family name, not a description of one fixed capability.

Think about an internal help tool. One version might read a policy PDF and draft an answer for an employee. Another could listen to a caller and speak back. A third might quietly search a library of manuals before the answering model writes anything. Those are different tasks, so they need different types of model behavior. Calling all of them “AI chat” hides the choices that affect what your application can accept and return.

AWS documentation also uses versioned guides. The Version 1 material describes separate understanding, creative and speech categories. The newer Nova 2 guide presents its current set around multimodal understanding, conversational speech, and multimodal embeddings. If you encounter an older Nova name in a tutorial, treat it as a clue to check the current model card and availability rather than proof that it is the right current option.

How a Nova model actually fits into an application

At a practical level, your application prepares an input and instructions, sends them through Amazon Bedrock, and receives the selected Nova model’s output. For a document assistant, the input might be a question plus retrieved passages from your company handbook. For a video workflow, it might be a clip and a request to identify the moments that need review. The model generates a result from patterns in its training and the supplied context; it does not turn into a verified source simply because the response sounds confident.

A model can be multimodal without doing every kind of generation. For example, a model that takes an image or a video and returns text is useful for describing, extracting, classifying, or answering questions about that material. It is not automatically an image or video generator. Likewise, a speech-to-speech model is designed around a conversational voice exchange, while an embeddings model returns vectors rather than a friendly written answer.

Some Nova 2 capabilities can add tools such as web grounding or code execution to a system. Those are system-level features that must be deliberately enabled and evaluated. They do not mean every Nova response is live, accurate, or independently verified. Your application still needs to decide which sources it accepts, what a tool is allowed to do, and how a user sees the result.

The model roles: understanding, voice, retrieval and older creative models

For current everyday multimodal work, AWS describes Nova 2 Lite as taking text, images, video and documents and producing text. That makes its role easy to picture: it can be used where a workflow needs a written result after looking at more than plain text, such as routing a document, extracting a field, summarising a clip, or preparing a first customer-support draft. The model choice is still a design decision, not a guarantee that every document or image will be read correctly.

For spoken interactions, the current Nova 2 guide lists Nova 2 Sonic as accepting speech and text, then returning speech and text. This is the kind of model role behind a voice agent that has to listen, keep track of a turn, and respond out loud. AWS’s live model materials use evolving version labels for its voice offering, so confirm the active model card before you build around a particular name, region, language, or connection limit.

Nova Multimodal Embeddings has a different job. It converts material into vectors: sets of numbers that allow software to find items with related meaning. If someone asks about a warranty, an embeddings system can help retrieve the most relevant passages before a text model drafts a response. It does not itself act like a chat assistant or generate an image. This distinction can prevent a frustrating mismatch between a search component and a reader-facing answer generator.

The older Version 1 guide is still helpful for understanding the wider Nova history. It categorizes Canvas as image creation and Reel as video creation, separate from understanding models and Sonic speech. That is useful context when you see those names, but it should not be read as a permanent, complete buying guide for today’s Nova 2 catalog. AWS updates its documentation and availability over time.

Nova vs Amazon Bedrock: the model and the platform

The shortest answer is that Nova is a model family, while Amazon Bedrock is the managed platform that gives an application access to foundation models. Bedrock is not limited to Nova: AWS says it supports models from Amazon and other providers. So choosing Bedrock does not by itself mean you have chosen Nova, and choosing Nova does not describe all the surrounding services that make an application run.

That surrounding platform can matter just as much as the model. Bedrock handles the request interface and offers related capabilities such as model access, security controls, logging options, evaluations, and other building blocks. An application team still owns the prompt, the data it sends, the permissions it configures, its retrieval setup, and the user experience. Saying “we use Nova” answers only the question of which intelligence family may process a request.

This is also why a web interface should not be confused with the model underneath. A chat screen or voice assistant is an application built around a model and platform. It may add its own tools, history, guardrails and data flows. When you compare products, ask whether you are comparing an interface, an API model, or the platform that hosts multiple models.

A practical situation to make the choice clearer

Hypothetical: a retailer wants an internal assistant for staff who need answers from product manuals, photographed labels and short training clips. A sensible first design question is not “Which AI is smartest?” It is whether the system must understand those input types and return a text answer. A multimodal text-output model may fit the answer step, while an embeddings model may organize the manual library so the system can retrieve relevant material first.

If the retailer later wants a phone-based staff helper, speech becomes a separate requirement. A speech-to-speech model can be evaluated for the conversation layer, but it still needs rules for when it may call an inventory tool or when it must pass a customer to a person. If the business wants product visuals or videos created, that is another requirement again; the older Nova documentation’s creative category shows why generation should be evaluated separately from understanding.

In this hypothetical system, Bedrock is the managed route through which the application calls the selected model or models. The model does not know the retailer’s policies unless the system supplies relevant information, and the system should test answers against representative materials before relying on it. Breaking the problem into input, output, retrieval, and oversight is more reliable than choosing a family name alone.

Limits, privacy context and responsible use

AI output is probabilistic. AWS explicitly says Amazon Nova can produce inaccurate or inappropriate content and recommends evaluating accuracy and suitability, especially when a result reaches end users. If an answer could affect a payment, employment decision, health action, legal position, or safety outcome, it needs risk assessment, testing, and meaningful human oversight. A fluent response is not enough evidence that it is correct.

Multimodal inputs create their own risks. A blurry photograph may be misread. A long document may contain contradictory instructions. A clip can omit the context needed to judge an event. Tool use can make a system more capable, but it can also broaden the consequences of a mistaken instruction. Define what the system is allowed to retrieve or do, test failure cases, and give users a route to correct or escalate an answer.

For Nova used through Amazon Bedrock, AWS states that Bedrock does not store or review customer prompts and completions, does not share them between customers or with Bedrock partners, and does not use them to train Bedrock models. That does not remove the need to review your own application’s logging, connected tools, retention choices, access controls, and applicable data rules. Treat the documented service statement as one part of the data path, not a blanket privacy promise for every implementation.

Frequently asked questions about Amazon Nova

These short answers cover the basic distinction. For a production decision, verify the active model documentation, region support, account access, and any feature constraints at the time you build.

The most useful first step is to name the input you have, the output you need, and the level of human review your situation requires. That turns a broad model-family question into an evaluable design choice.