Amazon Nova Forge: An 'Open Training' Paradigm That Lets Everyone Build Their Own Frontier AI

Amazon Nova Forge is a new service that helps organizations build expert foundation models tailored to their domain. The core idea is not a model that scores well on public benchmarks, but one that satisfies accuracy, reliability, cost efficiency, and control all at once in real work environments.

Previously, enterprises only had three paths to choose from.

• Fine-tuning a closed-weights LLM in a limited way, such as with LoRA
• Continuing pretraining/post-training on an open-weights model
• Building a frontier-scale model from scratch on their own

However, the first makes it hard to build deep domain understanding, the second carries a high risk of catastrophic forgetting since the original training data and recipes aren't available, and the third requires far too much compute and time.

To solve this problem, Nova Forge introduces a new approach called open training. Users are given checkpoints from the three stages—pretrained, mid-trained, and post-trained—along with recipes and code that let them continue training from each stage. On top of this, organizations can mix their own proprietary data with the curated data used to train Amazon Nova, allowing data and knowledge to be injected not just at one stage but across multiple stages.

The training stages break down as follows.

• pretraining: learning next-token prediction from large-scale unlabeled data
• mid-training: improving performance on specific tasks using real and synthetic user-system interactions (traces) while maintaining general-purpose ability
• post-training: strengthening instruction-following ability and policy alignment through SFT and RL

Thanks to this structure, for a completely new domain, one can continue pretraining from the pretrained checkpoint, and if sufficient historical data exists or interactions can be synthesized, one can continue from the mid-trained checkpoint. It's also possible to perform SFT and RL on the mid-trained checkpoint, or apply continuous RL updates using real feedback and synthetic data.

A particularly important feature is the API-based mixing capability that blends frontier-scale data with enterprise data. The authors explain that through this, domain expertise becomes a core capability of the model rather than an add-on feature, and the model handles both internal organizational knowledge and general knowledge with the same level of fluency.

As a case study, Nimbus Therapeutics is introduced. To accelerate molecular design and drug development, this company first used Nova 2 Lite for pharmaceutical patent analysis, achieving 95% accuracy without any separate customization. Later, aiming higher to build a unified molecular intelligence system, they applied Nova Forge, fed in their own chemistry datasets, and boosted performance using supervised fine-tuning and reinforcement learning.

In early results, this customized model significantly outperformed other leading LLMs on molecular property prediction tasks, and the potential to extend to molecule generation in the future was also suggested. This means the ability to find new drug candidates faster and speed up research and development, where experimentation costs are high.

Amazon describes Nova Forge as its first service that allows customers to build their own frontier models based on Amazon Nova. Going forward, they also plan to give customers early access to the more powerful Nova 2 Pro at the same time as their internal teams, and a few weeks later, to provide recipes that allow continued training from multiple checkpoints.

Ultimately, what Nova Forge aims for is reducing the cost of switching models. The goal is to shorten the time needed to transfer knowledge from a customized Nova to a newly released Nova model, and more broadly, to open a future where every organization can build its own frontier AI tailored to its domain.