Amazon Unveils New Frontier AI Models Along with Tools for Customers to Create Their Own

Amazon has introduced a new range of cutting-edge artificial intelligence models—and a novel option for users to create their own advanced models.

The ecommerce leader unveiled the second generation of its Nova AI models during re:Invent, a company event held in Las Vegas. While these models are not as widely recognized as those from competitors like OpenAI and Google, Amazon’s strategy to enhance their customizability could help them attract more cloud users.

Amazon presented two upgraded large language models, Nova Lite and Nova Pro, along with a new real-time voice model named Nova Sonic, and an experimental model referred to as Nova Omni, which simulates reasoning through images, audio, video, and text. These models are being rolled out today for a select group of customers.

More notably, considering the critical role of its cloud division, Amazon is launching a tool called Nova Forge that allows customers to create tailored frontier models by incorporating their own training data into incomplete versions of the Nova 2 Lite and Pro models.

Currently, it is possible to fine-tune existing AI models like Google’s Gemini and OpenAI’s GPT. However, Amazon’s method enables users to add data at different stages of the training process, including custom pretraining, typically reserved for large AI labs when building a base model.

“Everyone is looking for a frontier model that’s an expert in their domain,” Rohit Prasad, head of Amazon’s AI initiatives, shared with WIRED before the announcements. He explained that the technologies behind Nova Forge were developed to empower internal teams, including those working on Alexa and AI agents, to create bespoke models. “This represents a new open training paradigm,” he stated.

One user that has utilized this approach is Reddit, which employed Nova Forge to develop a specialized model for identifying content that violates community guidelines.

Conventional model fine-tuning wouldn’t suffice, according to Reddit’s chief technology officer Chris Slowe, because most models are programmed to entirely avoid offensive or violent content, thus refusing to analyze certain materials. Slowe explained that the combination of custom pre-training and traditional fine-tuning resulted in a model adept at understanding and engaging with Reddit.

“Other LLMs can grasp Reddit as a concept and understand how it functions, but they don’t delve into the specifics,” Slowe remarked. “We effectively created a model that is an expert on Reddit.”