A researcher reviews the Nova Forge training pipeline, where proprietary datasets blend with Nova-curated data to build customized frontier-scale models on AWS. Image Source: ChatGPT-5
Amazon Nova Forge Lets Enterprises Build Custom Frontier Models
Key Takeaways: Amazon Nova Forge
Build your own Nova-based frontier model, not just fine-tune one. Nova Forge gives customers access to early Nova checkpoints across pre-training, mid-training, and post-training, going deeper than typical fine-tuning or RLHF.
Data mixing is built in to reduce catastrophic forgetting. Organizations blend their proprietary datasets with Nova-curated data so the model can absorb new domain knowledge without losing core capabilities.
Reward functions run in the customer’s own environment. Teams can plug in their own scoring environments—like chemistry tools, robotics simulators, or complex agent workflows—for reinforcement learning that reflects real-world tasks.
End-to-end on AWS: Amazon SageMaker AI for training, Amazon Bedrock for serving. Models are trained on Amazon SageMaker AI infrastructure, then imported as private models on Amazon Bedrock, with consistent APIs and security controls.
Available today in one AWS Region. Nova Forge launches in US East (N. Virginia) and includes multiple Nova checkpoints, recommended training recipes, and integrations with both SageMaker AI and Amazon Bedrock.
Amazon Nova Forge: Custom Frontier Models for Your Own Domain
Organizations are racing to embed generative AI into every part of their business, but off-the-shelf models still struggle with deep domain expertise, proprietary workflows, and industry-specific language. Techniques like prompt engineering, RAG, and even post-training customization help, but they don’t rewrite a model’s core understanding of a domain—and fully training a frontier model from scratch is still out of reach for most companies in terms of data, compute, and cost.
Amazon Nova Forge aims to fill that gap. It’s a new service that lets organizations start their development from early Amazon Nova model checkpoints, mix their own proprietary data with Nova-curated training data, and then host their customized Nova-based frontier models securely on AWS. Amazon positions Nova Forge as the “easiest and most cost-effective” way to build your own Nova-based frontier-scale model while preserving the base model’s intelligence, instruction following, and safety behavior.
Why Nova Forge Exists: Limits of Today’s Customization
Most companies today adapt foundation models using prompt engineering, RAG with vector search, or later-stage techniques like supervised fine-tuning and reinforcement learning. These approaches work well for many applications, but they all sit on top of a fully trained model. That makes it harder to inject specialized knowledge deep into the model’s internal representation of a domain.
Some organizations try continued pre-training (CPT) with only their proprietary data. That can introduce a different problem: catastrophic forgetting, where the model’s original skills degrade as it learns new material. At the same time, training a modern frontier model from scratch requires huge datasets, large-scale distributed compute, and major capital outlays—still a non-starter for most enterprises.
Nova Forge is meant to sit in the middle: more control than fine-tuning, far less cost and complexity than building a model entirely from the ground up.
Use Cases: Who Nova Forge Is Designed For
Amazon says Nova Forge targets organizations with significant proprietary or industry-specific data that want models to “truly understand” their domain. Examples include:
Manufacturing and automation – Models that can reason about specialized processes, equipment logs, and plant-level workflows.
Research and development – Models trained on proprietary research, technical literature, or domain-specific methodologies.
Content and media – Systems that internalize brand voice, editorial standards, and custom moderation rules.
Specialized industries – Domains with dense, industry-specific terminology and regulation, where misinterpretation carries real risk (for example, finance, healthcare, or regulated utilities).
Across these scenarios, Nova Forge is pitched as a way to deliver:
more accurate task performance
differentiated capabilities tied to proprietary knowledge
improved latency and cost efficiency compared to generic models
Because customers start from Nova checkpoints rather than a blank slate, they inherit the base model’s general abilities and safety guardrails while layering in domain-specific behavior.
How Amazon Nova Forge Works
Training from Early Nova Checkpoints
With Nova Forge, customers choose from multiple Nova checkpoints—pre-trained, mid-trained, or post-trained—to match how deeply they want to participate in the model’s remaining training. They then:
Upload proprietary datasets or connect existing data sources in Amazon SageMaker Studio.
Blend those datasets with Nova-curated training data, which is organized by domain to help maintain general performance, instruction-following, and safety behavior, while reducing the risk of catastrophic forgetting.
Run training jobs using “proven recipes” on fully managed Amazon SageMaker AI infrastructure.
By starting from early Nova checkpoints and mixing proprietary data with Nova-curated datasets throughout training, customers can build highly specialized models without losing the broad reasoning, language understanding, and safety foundations that make frontier models useful in the first place.
Reinforcement Learning in Customer Environments
Nova Forge also supports reward-based training using customer-defined reward functions. That means teams can:
Run RL in environments that mirror their real use cases.
Use their own orchestrator to manage multi-turn rollouts, which is especially relevant for agents and sequential decision tasks.
Examples Amazon highlights include:
Chemistry tools that score molecular designs.
Robotics simulations that reward efficient task completion and penalize collisions.
By connecting these proprietary environments directly, organizations can push Nova-based models toward very specific behaviors without exposing internal IP to external systems.
Responsible AI Toolkit and Safety Controls
Nova Forge includes a built-in responsible AI toolkit that lets organizations configure safety and content moderation settings. Customers can adjust how the model handles:
sensitive or regulated content
security-relevant topics
business-specific guardrails
This allows different industries to tailor safety posture to their own policies while still benefiting from Nova’s baseline safety work.
From Training to Production: SageMaker AI + Amazon Bedrock
Nova Forge is tightly integrated into the existing AWS stack:
Training on Amazon SageMaker AI
Teams work from Amazon SageMaker Studio and use managed infrastructure to run large-scale training or continued pre-training.
Optional Reinforcement Fine-Tuning (RFT) can be used to improve factual accuracy and reduce hallucinations for specific domains.
Deployment via Amazon Bedrock
Once training completes, customers import the model into Amazon Bedrock as a private model.
Applications can then call this model using the same APIs they already use for other Bedrock models, with consistent security, monitoring, and integration across AWS services.
This end-to-end path is meant to lower friction: teams don’t have to design their own training clusters, stand up separate hosting infrastructure, or manage new endpoints. Instead, they stay within familiar AWS building blocks.
Amazon Nova Forge availability and enterprise support
Amazon Nova Forge is now available in the US East (N. Virginia) AWS Region. The service provides access to multiple Nova model checkpoints, AWS-designed training recipes for mixing proprietary and Nova-curated datasets, and seamless integration with Amazon SageMaker AI and Amazon Bedrock.
Organizations that want additional guidance can work directly with AWS’s Generative AI Innovation Center, which offers expert support for planning and executing custom model-development projects.
In Amazon SageMaker Studio, organizations can now build their own Nova-based frontier models using Amazon Nova.
Learn more in the Amazon Nova User Guide and explore Nova Forge from the Amazon SageMaker AI console.
Q&A: Amazon Nova Forge and Enterprise AI
Q: How is Nova Forge different from standard fine-tuning on Amazon Bedrock?
A: Traditional fine-tuning and RLHF typically operate on a fully trained base model. Amazon Nova Forge gives customers access to early Nova checkpoints and allows them to participate in deeper phases of training, blending their data with Nova-curated datasets throughout. The goal is a model whose core representation—not just its output polish—reflects the customer’s domain.
Q: How does Nova Forge address catastrophic forgetting?
A: Instead of training exclusively on customer data, Nova Forge uses data mixing strategies that combine proprietary datasets with Nova-curated data across all training phases. This approach is designed to preserve core intelligence, general instruction following, and safety behaviors while adding specialized knowledge.
Q: Where can customers use their custom models once they’re trained?
A: After training on Amazon SageMaker AI, customers can import their Nova-based models as private models in Amazon Bedrock. That lets them integrate the models into existing AWS applications using familiar APIs, security mechanisms, and monitoring tools.
Q: Where is Nova Forge available today?
A: At launch, Amazon Nova Forge is available in the US East (N. Virginia) AWS Region. The program includes multiple Nova checkpoints, training recipes, and integrations with SageMaker AI and Bedrock.
What This Means: Custom Frontier Models With Amazon Nova Forge
For most organizations, the hardest part of adopting generative AI isn’t spinning up an API—it’s getting models to genuinely understand their data, workflows, and risk boundaries. Amazon Nova Forge moves that conversation closer to the core of the Nova model family itself, while still keeping training and deployment inside a managed AWS environment.
If Nova Forge works as described, it could make frontier-class, domain-specific models a realistic option for enterprises that have strong data but don’t have the resources to build a model from scratch. The next challenge will be less about access to powerful models—and more about how responsibly, creatively, and rigorously organizations use this new level of customization.
Sources
Amazon – Nova Forge Announcement
https://aws.amazon.com/blogs/aws/introducing-amazon-nova-forge-build-your-own-frontier-models-using-nova/
Amazon – Nova Model Family Overview
https://aws.amazon.com/nova/
Amazon – Prompt Engineering Overview
https://aws.amazon.com/what-is/prompt-engineering/
Amazon – Retrieval-Augmented Generation (RAG)
https://aws.amazon.com/what-is/retrieval-augmented-generation/
Amazon SageMaker – Continued Pre-Training (CPT) for Nova
https://docs.aws.amazon.com/sagemaker/latest/dg/nova-cpt.html
Amazon – Reinforcement Learning Overview
https://aws.amazon.com/what-is/reinforcement-learning/
Amazon – Global Infrastructure (Regions & Availability Zones)
https://aws.amazon.com/about-aws/global-infrastructure/regions_az/
Amazon SageMaker AI
https://aws.amazon.com/sagemaker/ai/
Amazon Bedrock Overview
https://aws.amazon.com/bedrock/
Amazon SageMaker Studio
https://aws.amazon.com/sagemaker/ai/studio/
Amazon SageMaker Console
https://console.aws.amazon.com/sagemaker/home
Amazon Nova User Guide
https://docs.aws.amazon.com/nova/latest/userguide/what-is-nova.html
AWS Generative AI Innovation Center
https://aws.amazon.com/ai/generative-ai/innovation-center/
Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.