Amazon Nova Forge and Open Training for Domain Specific AI Models

Introduction

Organizations are scaling generative AI across functions from customer support and analytics to engineering and operations. But as soon as you move from generic tasks to deep domain work (your proprietary processes, terminology, and data), the usual customization playbook starts to creak. Prompt engineering and RAG can be powerful, yet they don’t truly “bake” specialized knowledge into the model. Traditional fine-tuning helps, but it often happens late in the lifecycle, after the base model is already heavily shaped, making it harder to steer toward a niche domain without trade-offs.

Amazon Nova Forge is AWS’s answer to that gap: a service designed to help teams build their own frontier-class models by starting earlier in the training journey, blending their data with Nova-curated datasets, and then deploying the result securely on AWS.

Why “normal” customization hits a ceiling?

Most teams begin with RAG because it’s fast and low-risk: you keep the model fixed and feed it relevant context at inference time. But RAG can struggle when the task demands embedded expertise, consistent reasoning, and proprietary workflows, internal jargon, or nuanced standards that aren’t easily retrieved as standalone snippets. Meanwhile, continued pre-training on only proprietary data can lead to catastrophic forgetting, where the model improves on your niche content but loses general capabilities that made it useful in the first place.

Training from scratch is an option in theory, but the cost, data scale, and ML expertise required make it prohibitive for most organizations.

Amazon Nova Forge

Amazon Nova Forge introduces an “open training” approach: you get access to early Nova checkpoints across major stages of model development, plus AWS-provided recipes to mix your proprietary data with Nova-curated training data across training phases. This is meant to maximize how much the model learns from your domain while preserving foundational skills like reasoning and instruction-following.

In practical terms, Amazon Nova Forge is positioned as a guided path to building a “private edition” of a Nova model trained for your business and hosted on AWS.

The core idea: start from the right checkpoint

Nova Forge supports development from checkpoints spanning:

• Pre-training (high plasticity, best for absorbing large volumes of domain data)
• Mid-training (more targeted adaptation with conservative learning dynamics)
• Post-training (alignment stages like supervised fine-tuning and reinforcement learning)

Data mixing: the antidote to catastrophic forgetting

A major differentiator in Amazon Nova Forge is data mixing, blending your proprietary corpus with Nova-curated data during training. Instead of teaching the model on a narrow distribution (risking instability and forgetting), mixing keeps the model anchored to broad, frontier-scale patterns while it absorbs your specialty knowledge.

Reinforcement Fine-Tuning and responsible AI controls

Beyond supervised learning, Amazon Nova Forge supports reinforcement learning, where you can bring reward functions from your own environment, useful when correctness depends on business-specific scoring (e.g., simulations, tools, evaluation harnesses, multi-step agent workflows). The AWS announcement highlights RL as a way to learn from feedback produced in environments representative of real use cases.

Amazon Nova Forge also includes a built-in responsible AI toolkit to configure safety and content moderation settings for your custom model, letting you tune guardrails to your organization’s policies and risk posture.

How the end-to-end workflow fits into AWS?

A typical Amazon Nova Forge journey looks like this:

1. Choose a starting checkpoint (pre-trained, mid-trained, or post-trained) based on how much data you have and what you’re optimizing for.
2. Train on AWS managed infrastructure using Amazon SageMaker AI (with proven recipes and workflows
3. Import and serve the resulting model in Amazon Bedrock as a private model, using consistent APIs and security controls similar to other Bedrock-hosted models
4. This pairing of Amazon SageMaker for training and Bedrock for deployment targets teams that want customization depth without building an entire frontier-model platform from scratch.

Who is Amazon Nova Forge for?

AWS positions Amazon Nova Forge for organizations that have meaningful proprietary or industry data and want models that truly internalize domain context, such as manufacturing workflows, R&D corpora, brand-specific content rules, and regulated or specialized industries.

If your use case is mostly “retrieve the right policy paragraph and summarize it,” RAG might be enough. But if you need the model to reason like an insider, make consistent decisions, produce structured outputs, use domain-native language, and follow tool-driven workflows, Amazon Nova Forge aims to move domain expertise from “context you attach” to “capability the model has.”

Availability and getting started

The launch announcement notes Nova Forge availability in US East (N. Virginia), with a workflow that begins in Amazon SageMaker (including SageMaker Studio experiences) and ends with private model hosting in Amazon Bedrock.
Amazon’s SageMaker developer guide also describes subscription prerequisites and access steps for Amazon Nova Forge.

Conclusion

Amazon Nova Forge is AWS’s bet that the next wave of enterprise GenAI will be driven by open training: letting customers start from earlier training checkpoints and blend proprietary data with curated frontier data to build truly domain-smart models, then deploy them with enterprise-grade controls on AWS.

Drop a query if you have any questions regarding Amazon Nova Forge and we will get back to you quickly.

About CloudThat