Amazon Web Services Inc., the cloud division of Amazon.com Inc., today announced a new family of multimodal, generative artificial intelligence models called Nova.

Amazon Chief Executive Andy Jassy debuted the new family of models onstage at the AWS re:Invent conference in Las Vegas. It includes four models focused on text outputs named Micro, Lite, Pro and Premier. Each of them represents a model of increasing size, complexity and capability. The first three are immediately available, whereas the most advanced, Premier, is still in training and will arrive in early 2025.

Alongside these four models, Amazon also released two creative models: Nova Canvas, an image generator model, and Nova Reel, a video creation model.

Micro, the smallest large language model, is text-only and provides the fastest responses for a very low cost. The model itself was designed assist with text summarization, translation, question and answer, conversational chat and brainstorming.

Lite is the next step up providing low-cost multimodal support for processing images, video and text inputs to generate text responses. Amazon said that model will be worthwhile for real-time customer interactions and document analysis where visuals would be involved. The model can process 300,000 tokens in input, which is around the length of three ordinary novels combined, analyze multiple images at once or up to 30 minutes of video in a single request.

Pro, currently the most capable multimodal LLM available from AWS, combines all of the capabilities of the previous models and sets high standards for AI agents. Agents are a type of AI capability that can take action on the behalf of humans without supervision and use third-party tools in order to complete complex activities. For example, Pro could be used to write and send emails, or gather data, complete reports and distribute them without the need for additional external action.

According to Amazon, the Pro model can also act as a “teacher” model to help create custom variants of Nova Micro and Lite. Larger, more capable models are often used to act as a source of knowledge to help fine-tune less complex, more efficient “student” models. This allows the smaller model to achieve similar performance but use less computational power and memory.

The company stressed that what makes Nova particularly useful is the ability to tailor it to enterprise and industry needs. Foundation models are designed to act as a starting point that can be fine-tuned and adjusted to understand an industry’s particular terminology, fit a brand voice and optimize on enterprise data. For example, a healthcare firm might fine-tune Amazon Nova to understand medical terminology, forms and relationships in the industry.

Nova Canvas offers state-of-the-art image generation that can create professional images from text prompts or images provided to it. Users can also edit images using text inputs, including identifying objects or spaces in the image that the user wants to change. The user needs only state something such as “shirt,” and then provide an English prompt of what they want on the shirt visible in the image and Canvas will change the contents of the shirt to match.

The user can also ask Canvas to maintain, or change, backgrounds and color schemes according to user preference. Everything about an original or edited image can be modified according to prompts.

Reel produces short videos from text prompts similar to other high-fidelity text-to-video AI models on the market. Prompts can include natural language describing camera motion such as zoom, side-to-side and rotation, which allows the user to easily create cinematic shots.

The Amazon Nova text generating models understand and generate content in over 200 languages, with powerful capabilities in English, German, Spanish, French, Italian, Japanese, Korean, Arabic, Simplified Chinese and Russian. The creative models, Canvas and Reel, support only English prompts.

The new Nova models, except Premier, are available today on Amazon Bedrock, an AWS managed service that provides access to cloud-hosted frontier AI models from Amazon and other providers, along with a set of tools for building AI applications.

Photo: Robert Hof/SiliconANGLE