AWS Launches New Nova AI Models and a Service That Gives Customers More Control

Amazon Web Services is introducing a new lineup of proprietary AI models, along with a service designed to help enterprise customers create their own customized versions.

During AWS CEO Matt Garman’s keynote at AWS re: Invent on Tuesday, the company unveiled Nova 2, a group of four upgraded AI models added to the Nova family.

The original Nova lineup debuted last year at AWS’s annual tech conference, when the company released four text-generation models and one image-generation model. This year’s launch delivers major upgrades, along with a companion service for deeper customer customization.

“The momentum has been really fantastic,” Garman said during the keynote. “Nova has grown to be used by tens of thousands of customers today — everyone from marketing giants to tech leaders like Infosys, Blue Origin, or Robinhood, to innovative startups like NinjaTech AI — and today we’re making Nova even better.”

The four new models include:

• Nova 2 Lite — a lower-cost reasoning model capable of processing text, images, and video to produce text responses for everyday tasks.
• Nova 2 Pro — an advanced reasoning agent designed for “highly complex tasks,” including coding, with support for text, images, video, and speech.
• Nova 2 Sonic — a speech-to-speech model built for conversational AI experiences.
• Nova 2 Omni — a multimodal reasoning and generative model that accepts images, text, video, and speech input and generates both text and pictures.

In addition to the model enhancements, AWS introduced a new service called Nova Forge, which allows cloud customers to develop their own frontier-style versions of the Nova models — known as Novellas — for $100,000 per year, per CNBC. The service provides access to models at pre-trained, mid-trained, or post-trained stages, enabling companies to train them with their proprietary data further.

Garman said the offering is meant to resolve challenges that arise when companies try to adapt general-purpose AI models with large amounts of internal data.

“The more you customize models, the more you add data in post-training, these models tend to forget some of the interesting stuff they learned earlier — the core reasoning,” Garman explained. “It’s a bit like humans trying to learn a new language. When you start young, it comes easily, but learning it later in life is much harder. Model training works similarly.”

Early adopters of Nova Forge include Reddit, Sony, and Booking.com.

Readers can follow TechCrunch’s complete coverage of the annual enterprise tech conference for more updates.