Amazon Nova Forge Lets Enterprises Build Custom Frontier Models Without Starting From Scratch

Amazon has launched Nova Forge, a new service that lets enterprises build their own frontier AI models using the Nova family of foundation models as a starting point. The offering addresses a fundamental challenge in enterprise AI: how to embed specialized domain knowledge into AI systems without sacrificing the core capabilities that make them useful in the first place.

The Core Problem Nova Forge Solves

Organizations across industries are racing to deploy generative AI in every corner of their business. But when applications require deep domain expertise or specific business context, standard foundation models often fall short. They lack the proprietary knowledge, industry-specific terminology, and specialized workflows that make AI genuinely useful for enterprises.

Two common approaches have dominated the conversation: prompt engineering and Retrieval Augmented Generation (RAG). Both work reasonably well for many use cases, but they have inherent limitations. They operate on top of a fully trained model, meaning the model's core understanding never truly absorbs specialized knowledge. The domain expertise lives in a retrieval layer or prompt context, not in the model's fundamental reasoning.

The alternative has been supervised fine-tuning or continued pre-training with proprietary data. But this path is fraught with problems. When organizations attempt what AWS calls Continued Pre-Training (CPT) using only their proprietary data, they frequently encounter catastrophic forgetting—where models essentially lose their foundational capabilities as they learn new content.

The model becomes excellent at domain-specific tasks but terrible at everything else.

At the same time, the resources required to train a frontier model from scratch place this capability out of reach for most organizations. The data, compute, and cost barriers are simply too high.

How Nova Forge Changes the Equation

Nova Forge tackles these problems by allowing enterprises to start model development from early checkpoints across pre-training, mid-training, and post-training phases. Instead of receiving a fully formed model, customers choose from Nova checkpoints at different stages of the training pipeline.

Checkpoint Selection

• Pre-training checkpoint: Offers maximum plasticity for ingesting large volumes of domain data
• Mid-training checkpoint: Balances plasticity with retained general knowledge
• Post-training checkpoint: After alignment through instruction tuning and RLHF, suitable for fine-tuning on niche tasks

The key innovation is what AWS calls data mixing. Throughout all training phases, customers can blend their proprietary data with Amazon-curated Nova training data. This approach significantly reduces catastrophic forgetting compared to training with raw proprietary data alone, helping preserve foundational skills including core intelligence, instruction following capabilities, and safety guardrails.

Research cited by Amazon Science demonstrates this works in practice. When using data mixing, models retain MMLU benchmark performance within 0.01 of baseline while boosting domain-specific F1 scores by 12 points. That's a meaningful trade-off for enterprises that need domain expertise without destroying general capability.

Beyond Basic Training: Reinforcement Fine-Tuning

Nova Forge also provides reinforcement learning capabilities that let customers use reward functions in their own environments. This allows the model to learn from feedback generated in environments representative of specific use cases.

For complex agent workflows and sequential decision-making tasks, customers can use their own orchestrator to manage multi-turn rollouts. Whether the environment involves chemistry tools scoring molecular designs or robotics simulations that reward efficient task completion, enterprises can connect proprietary environments directly to the training pipeline.

This is a meaningful departure from simple fine-tuning. The model isn't just learning from labeled examples; it's learning from feedback loops that reflect real-world domain-specific success criteria.

Enterprise Use Cases

Nova Forge targets organizations with proprietary or industry-specific data who want AI that genuinely understands their domain. The service outlines several key use cases:

• Manufacturing and automation: Build models that understand specialized processes, equipment data, and industry-specific workflows—rather than forcing workers to adapt to generic AI interfaces
• Research and development: Embed proprietary research data and domain-specific knowledge in the model itself, going beyond RAG-style retrieval to genuine understanding of scientific contexts
• Content and media: Develop models that understand brand voice, content standards, and specific moderation requirements absorbed into core reasoning rather than requiring elaborate prompt engineering
• Specialized industries: Train on industry-specific terminology, regulations, and best practices—healthcare compliance, financial services requirements, or legal documentation standards

Integration with the AWS Ecosystem

Nova Forge integrates with existing AWS workflows, which is critical for enterprise adoption. Customers run training using Amazon SageMaker AI's managed infrastructure, then import custom Nova models as private models on Amazon Bedrock.

This brings major practical benefits. The custom models get the same security protections, consistent APIs, and broader AWS integrations as any model in Amazon Bedrock. Organizations don't need to build separate inference infrastructure or manage their own model hosting.

The service is currently available in the US East (N. Virginia) AWS Region, with access to multiple Nova model checkpoints, training recipes, and integration with SageMaker and Bedrock.

Implications for the Enterprise AI Landscape

Nova Forge represents a meaningful shift in how enterprises can approach custom AI. Rather than choosing between generic foundation models that don't understand their domain or fine-tuned models that have lost general capabilities, organizations now have a third path.

The economic argument is also worth considering. Training a frontier model from scratch requires enormous investment in data, compute, and expertise. Nova Forge provides a middle ground: domain-specific customization at a fraction of the cost, with built-in mechanisms to preserve what makes the base model useful.

For enterprises with significant proprietary data assets, this could be transformative. Companies with unique data have long been told that AI is their competitive advantage, but turning that data into genuinely capable AI required either dangerous compromise on model quality or prohibitively expensive custom training. Nova Forge changes that calculation.

The broader question is how quickly organizations will adopt this approach. Early customers reportedly include Reddit, Booking.com, and Sony, building domain-specific models they call "Novellas." If these deployments prove successful, expect rapid acceleration in enterprise interest.

The availability in US East suggests this is a limited rollout, with broader geographic expansion likely following standard AWS patterns. Organizations outside this region will need to weigh whether immediate adoption justifies cross-region infrastructure costs, or whether waiting for regional availability makes more sense strategically.