A new Amazon service lets business customers build their own frontier models off of the Bedrock AWS managed services hub.
Billed as an “open training” service, this new application of Amazon Nova models lets the user select from various choices to train a custom large language model, blending proprietary data with Amazon’s Nova data sets for a unique “novella” – a domain-specialized frontier model that can be used for all kinds of exciting enterprise tasks. Reinforcement fine-tuning and other features are also available.
As of press time, this high-end enterprise offering may be currently only available in Northern Virginia, but you can bet that Amazon is in the process of rolling it out elsewhere, and letting a wider spectrum of business customers use it.
Nova Forge and Model Selection
One of the big features of Amazon Nova Forge is that business users can choose pre-trained, mid-trained or post-trained models.
What that means is that you can tap into the process at a different stage of model development.
For a pre-trained model, there’s more of a focus on generic data, and there’s less locked into the process at the early stage.
A mid-trained model has learned a lot, but still has a significant way to go, where a post-trained model is mostly done with things like instruction tuning, and ready for domain adaptation.
Who’s Using Nova Forge?
Available reporting names two businesses that are currently using Amazon Nova Forge models.
One is Sony, which is integrating Amazon Nova Forge with its agentic AI structure for fan engagement and other uses.
A press release explains more about how this is done over at the long-renowned sound tech brand:
“Sony is harnessing AWS’s comprehensive set of services to empower its global employees to innovate faster and operate more efficiently, while also creating a deeper connection and engagement between fans and creators across all its operating companies.”
There’s more, including details on not just what specific things Nova Forge is empowering at Sony, but how many task instances it processes on a regular basis:
“Sony’s internal enterprise AI platform gives employees across its entire group of companies access to generative AI and agentic AI services. Amazon Bedrock AgentCore helps Sony build, deploy, and manage AI agents seamlessly, while providing enterprise-grade security, observability, and scalability across the entire organization. The platform currently processes 150,000 inference requests daily and is expected to grow 300-fold in a few years, helping employees draft content, respond to inquiries, forecast, detect fraud, brainstorm, and develop new ideas.”
The other business making use of Amazon Nova Forge right now is Reddit, that popular social commentary platform that is reportedly training Nova custom models on its own platform data.
I also found this obscure list of other current customers from SiliconAngle, where writer Kyt Dotson calls Nova Forge a “bespoke” flavor of AI: Booking.com B.V., Cosine AI, Nimbus Therapeutics, Nomura Research Institute, and OpenBabylon.
And what do these firms do with the tech? Press resources cite applications like content and media, scientific research and development, and manufacturing automation.
Open, Closed and Custom Models
Can Nova Forge help “solve” the closed vs. open model dilemma, where companies have to weigh the benefits and risks of vendor services or internal ownership?
Well, first, to clarify, with Nova Forge, you get access to training phases and recipes, not to the raw source code or public weights.
Customers can run domain-specialized tasks on the private Bedrock model, giving the customer more control over behavior, evals, and safety policies than you’d get with a generic public API. On the other hand, critics point out, the end product, the “Novella,” is still tied to AWS infrastructure and APIs, so technically, one could call it platform lock-in, not open infrastructure. And again, it’s only partly open.
But Amazon is probably betting it can allay customer concerns: this is a multi-billion-dollar project as Amazon competes with Microsoft and Google for cloud customers.
It’s easy to imagine the direct appeal that this has to companies that have struggled with “build vs. buy” as AI systems quickly become more capable. Agentic AI requires quick thinking, and for some, this is the solution of its time. For others, there are other “flavors” of AI, either vendor-supported or in-house, that will bring companies confidently forward into the light, as we head into 2026.
