2 min Analytics

Amazon Nova models can process text, photos and videos

Amazon Nova models can process text, photos and videos

Nova is the new family of Amazon multimodal models intended to excel in adaptability and better suit a company’s needs. This type of model can handle both textual and visual input.

AWS is making six different models available for this purpose. Four of them—Micro, Lite, Pro, and Premier— generate textual output after analyzing a prompt. The differences are in size, complexity, and capabilities. In addition, there are the visual models Nova Canvas and Nova Reel, which generate images and videos, respectively.

The customizable nature means that industry-specific terms can be included, the language of the business can be spoken, and the model can be optimized for enterprise data. Something like this can be useful for law firms and hospitals, for example, where people use many specific terms and there are additional requirements around data.

Differences

As the name suggests, Nova Micro is the smallest large language model. It is built for generating summaries, translations, answers to questions, conversations, and brainstorming. Companies expecting a responsive and less expensive model may choose Nova Micro.

Amazon also describes Lite as a low-cost multimodal model. This option is especially suitable for companies seeking real-time visual customer interactions and document analysis support. The model can process 300,000 tokens of input, analyzing multiple images simultaneously or handling 30 minutes of video in a single request.

Nova Pro goes further than Micro and Lite by also autonomously performing further actions. For example, it can write and send emails and complete and distribute reports. Nova Pro can also function as a “teacher” for Micro and Lite. That is, the larger Pro model acts as a source of knowledge for fine-tuning the two Nova brothers. Thus, the smaller models can still perform similarly but use less computing and memory than the Pro version.

Micro, Lite, and Pro are all available immediately, as are the more visual models Canvas and Reel. However, there is limited availability for the US East (N. Virginia) region for now. Amazon is still developing Premier and aims for a release early next year.

Tip: AWS introduces multi-agent orchestration for Bedrock