Amazon announced a modern family of pioneering AI models and a modern way for customers to create their own pioneering models.
The e-commerce giant announced the second generation of its Nova AI models at re:Invent, a corporate conference held in Las Vegas. These models aren’t as popular as those offered by competitors like OpenAI and Google, but Amazon’s plan to make them highly customizable could make them popular among cloud users.
Amazon detailed two upgraded enormous tongue models, the Nova Lite and Nova Pro; a modern real-time voice model called Nova Sonic; and a more experimental model called Nova Omni, which performs a simulated type of reasoning using images, audio and video, as well as text. The modern models are made available to a restricted number of customers today.
More importantly, given the importance of its cloud business, Amazon is also releasing a tool called Nova Forge that will allow customers to create specialized boundary models by adding their own training data to unfinished versions of Nova 2 Lite and Pro models.
It is already possible to fine-tune off-the-shelf AI models such as Google’s Gemini and OpenAI’s GPT. However, Amazon’s approach allows customers to add data at various stages of model training, including the base model building process, a stage called custom pre-training that is typically reserved for enormous AI labs.
“Everyone is looking for a pioneering model who is an expert in their field,” Rohit Prasad, who leads Amazon’s artificial intelligence efforts, told WIRED ahead of today’s announcements. Prasad says Amazon developed the technologies behind Nova Forge to enable internal teams, including those developing Alexa and AI agents, to create custom models. “It’s a fundamentally new paradigm for open training,” he says.
One client that has already tested this approach is Reddit, which used Nova Forge to create a custom model to identify content that violates the platform’s rules.
Refining a conventional model wouldn’t work, says Reddit’s chief technology officer Chris Slowe, because most models are designed to avoid offensive or violent content altogether, meaning they refuse to analyze certain material. Slowe says custom pre-training combined with conventional tuning has created a pioneering model that is expert in understanding and using Reddit.
“Other LLM companies understand Reddit as a concept and how it works, but don’t dive into the details,” Slowe says. “We really built the Reddit expert model.”