AWS says anyone can build an AI model with Amazon Nova Forge

AWS has unveiled a new service aimed at helping enterprises build AI models in a bid to democratize access.

Amazon Nova Forge, unveiled at the company’s annual re:Invent conference in Las Vegas, will allow customers to “blend” their own in-house data with the company’s Nova model range.

Building a frontier AI model from scratch is a costly process, and one that most businesses simply can’t afford, CEO Matt Garman told attendees at the conference. In launching the service and combining customer data with its own open weight models, this bridges the gap for those without the necessary financial resources.

“Nova Forge is a new service that introduces the concept of open training models,” he said.

“With Nova Forge, you get exclusive access to a variety of Nova training checkpoints, and then you get the ability for you to blend in your own proprietary data together with an Amazon-curated training dataset at every stage of the model training.”

These “checkpoints” cited by Garman refer to stages within the training process of an AI model, whether that be pre-trained, mid-trained, and post-trained. Users can integrate their own data at certain points of this process depending on their individual requirements.

The advantage here, Garman claimed, is that the service both reduces the technical and financial load on the enterprise while giving them a bespoke model that “deeply understands” their data and specific business needs or area of expertise.

This understanding of the business’ knowledge base also has long-term advantages, allowing for easier fine-tuning and upgrades down the line.

“Because your baseline model already understands your business, these post-training techniques are actually much more effective,” he said. “Once you’re ready, you import this model, your Novella, into Bedrock, and you run inference on it just like you would any other Bedrock model.”

Customers will be able to begin building their own Novella custom models using its Nova 2 Lite model, also announced at the conference yesterday.

AWS bets big on Nova Forge

The announcement from AWS underlines the company’s long-term focus in terms of AI development and adoption: flexibility and customer choice.

As Garman noted, those who build models using the Forge service will be able to deploy it on Amazon Bedrock, the company’s platform which gives customers access to a range of third-party and in-house AI models.

This platform has been the foundation upon which AWS has built out its AI services over the last two years. In recent months, the company has added new agentic AI-related options through the Bedrock AgentCore service.

Keeping these models within the AWS ecosystem has added benefits, the company noted, particularly in terms of security and scalability.

“This complete solution – from building their own frontier model to production deployment – ensures organizations achieve optimal AI performance tailored to their specific business needs, with exclusive use of their model security hosted on AWS,” the hyperscaler said in a statement.