Amazon Has New Frontier AI Models—and a Way for Customers to Build Their Own

Amazon has announced a new family of frontier artificial intelligence models—and a new way for customers to build frontier models of their own.

The ecommerce giant announced the second generation of its Nova AI models at re:Invent, a company conference held in Las Vegas. The models are nowhere near as popular as those offered by rivals like OpenAI and Google, but Amazon’s plan to make them highly customizable could see them gain traction with its cloud users.

Amazon detailed two improved large language models, Nova Lite and Nova Pro; a new realtime voice model called Nova Sonic; and a more experimental model called Nova Omni that performs a simulated kind of reasoning using images, audio, and video as well as text. The new models are being made available today to a limited number of customers.

More significantly, given the importance of its cloud business, Amazon is also releasing a tool called Nova Forge that will let customers create specialized frontier models by adding their own training data to unfinished versions of the Nova 2 Lite and Pro models.

It is already possible to fine-tune off-the-shelf AI models like Google’s Gemini and OpenAI’s GPT. But Amazon’s approach lets customers add data at various stages of model training, including the process of building the base model, a stage known as custom pre-training that is normally reserved for large AI labs.

“Everyone is looking for a frontier model that’s an expert in their domain,” Rohit Prasad, who leads Amazon’s AI efforts, told WIRED ahead of today’s announcements. Prasad says that Amazon developed the technologies behind Nova Forge to empower internal teams, including those developing Alexa and AI agents to build custom models. “This is essentially a new open training paradigm,” he says.

One customer that has already tested the approach is Reddit, which used Nova Forge to create a custom model to identify content that breaks the platform’s rules.

Fine-tuning a conventional model would not work, says Reddit chief technology officer Chris Slowe, because most models are designed to avoid offensive or violent content entirely, meaning they would refuse to analyze some materials. Slowe says that custom pre-training, combined with conventional fine-tuning, produced a frontier model that is expert at understanding and using Reddit.

“Other LLMs understand Reddit as a concept, and how Reddit works, but they’re not down in the weeds,” Slowe says. “We really built a Reddit expert model.”