AWS re:Invent 2025 - Amazon Nova Forge: Build your own frontier models using Amazon Nova (AIM3325)

Democratizing Frontier AI with AWS Nova Forge

Overview

• The session introduced AWS Nova Forge, a new capability that allows organizations to build their own frontier AI models by leveraging the power of the AWS Nova family of foundation models.
• The key focus was on how Nova Forge enables companies to bridge the gap between their own intellectual property/domain knowledge and the capabilities of existing foundation models.

Foundation Model Construction

• The presenters explained the typical 3-stage process of building a foundation model:
  1. Pre-training: Leveraging large amounts of general web/text data to build broad world knowledge.
  2. Mid-training: Further enhancing the model's capabilities in a specific domain using more targeted data.
  3. Fine-tuning: Refining the model's performance on the organization's specific use cases and data.
• This process was analogized to dog training, where the initial obedience training is followed by reinforcement learning in real-world environments.

Key Benefits of Nova Forge

1. Access to Checkpoints: Nova Forge provides access to model checkpoints at each stage of the training process, allowing organizations to start at the appropriate point based on their data.
2. Blending of Data: Nova Forge enables seamless blending of an organization's proprietary data with the curated data used to train the Nova models, mitigating the risk of "catastrophic forgetting".
3. Optimized Recipes: Nova Forge offers SageMaker Hyperparameter Tuning recipes that are pre-optimized to maximize performance when using Nova as the starting point.
4. Reinforcement Learning: Nova Forge allows organizations to bring their own environments and orchestrators to the training process, enabling more advanced reinforcement learning approaches.
5. Safety and Responsibility: Nova Forge provides access to the same responsible AI controls and safety measures used in training the Nova models, allowing organizations to adapt them to their specific needs.

Real-World Examples

1. Reddit: Used Nova Forge to build a specialized content moderation model that achieved a 25% reduction in missed threats, while also improving precision and recall.
2. Nimbus Therapeutics: Leveraged Nova Forge to build a drug discovery assistant model that outperformed specialized graph neural network models, while benefiting from lower latency and cost.
3. Numero Research Institute: Developed a financial services language model for the Japanese market that exceeded the performance of models built using open-source approaches.
4. Sony Group: Created a legal research assistant model using reinforcement learning on Nova Forge, surpassing the performance of larger language models.

Key Takeaways

• Nova Forge democratizes access to frontier AI by allowing organizations to build custom foundation models tailored to their specific needs and data.
• The ability to leverage checkpoints, blend data, and utilize optimized recipes significantly reduces the time and cost of building high-performing foundation models.
• Reinforcement learning capabilities enable organizations to integrate their own environments and workflows, unlocking more advanced use cases.
• The safety and responsibility features ensure the models can be adapted to meet an organization's specific compliance and ethical requirements.
• Real-world examples across industries demonstrate the tangible benefits of using Nova Forge to build custom frontier AI models.