Amazon’s Rohit Prasa outlines “open training” approach behind Nova Forge launch

Rohit Prasad, senior vice president and head scientist for Artificial General Intelligence at Amazon, has shared new technical detail on Amazon Nova Forge, outlining how its open training approach allows organizations to build domain-specific foundation models without losing core capabilities.

Amazon has provided further insight into the technical foundations of Amazon Nova Forge, with Rohit Prasad, senior vice president and head scientist for Artificial General Intelligence at Amazon, explaining how the service enables organizations to train frontier-scale AI models using proprietary data while avoiding what the industry calls “catastrophic forgetting.”

The update highlights a broader shift toward customizable foundation models as enterprises, including education and skills providers, seek AI systems that perform reliably beyond public benchmarks.

Why Amazon built Nova Forge

In a LinkedIn post, Prasad explains that Nova Forge was initially developed to meet internal demand across Amazon’s businesses, where teams required models with deep expertise in specific domains rather than general-purpose performance.

“We built Forge initially because our own internal teams needed it,” Prasad says. “Across Amazon's diverse businesses, teams needed models with deep expertise in their specific domains – and our external customers wanted the same.”

According to Prasad, two core challenges shaped the development of Nova Forge. The first was how to integrate large volumes of domain-specific data into a model without degrading its foundational capabilities. The second was how to make advanced training tools and workflows accessible to customers without requiring specialist AI research teams.

These challenges reflect a common issue across enterprise AI deployments, where models that perform well in benchmarks often struggle when applied to organization-specific data and evolving real-world contexts.

Open training and model checkpoints

Amazon describes Nova Forge as being built around an open training paradigm. Rather than relying solely on post-training fine-tuning, the service provides access to model checkpoints across three stages of development: pretraining, mid-training, and post-training.

This allows organizations to continue training models at the stage most relevant to their use case. Teams working in novel or highly specialized domains can introduce proprietary data earlier in the training process, while others may focus on supervised fine-tuning or reinforcement learning at later stages.

Prasad says the approach required months of science and engineering work, including refining training recipes and validating them with internal teams and early customers on Amazon SageMaker AI.

“We spent months refining these recipes with internal teams, then validated them with early customers to make each training stage accessible and predictable,” Prasad says. “Each time a customer used Forge recipes, their model significantly outperformed other leading alternatives.”

A central feature of the platform is the ability to mix proprietary data with the curated frontier-scale data used to train Amazon Nova models at every stage, which Amazon says helps prevent regression in core model performance.

How Nova Forge is positioned for enterprise use

Amazon frames Nova Forge as a response to a growing gap between benchmark performance and production reality, where foundation models often struggle once deployed in organization-specific environments. Models trained primarily on public datasets can fail to reflect proprietary data, internal workflows, and changing user behavior.

By allowing domain expertise to be embedded during core training stages, Amazon says Nova Forge enables organizations to build models where specialized knowledge is a foundational capability rather than an added layer. The company positions this approach as particularly relevant for sectors working with complex or sensitive data, including healthcare, scientific research, and applied learning and skills development.

Amazon also confirms that Forge customers will receive early access to newer models, including Nova 2 Pro, alongside updated training recipes designed to support knowledge transfer between model generations.

“This is just the beginning of what's possible when we lower the barriers to frontier model development,” Prasad says. “Excited to see what organizations build through this ‘open training’ paradigm.”