AWS Nova Forge: Open-Training Platform Enables Deep Customization of Nova 2 Models
Ddos 
AWS first unveiled its homegrown Nova series models at last year’s re:Invent, and has since continued expanding the lineup. At re:Invent 2025, the company introduced Nova 2 Lite, Nova 2 Pro, and Nova 2 Sonic, alongside Nova 2 Omni—a single model capable of executing multimodal tasks—as well as Nova Act, designed to simulate human interactions within a browser, and Nova Forge, a framework enabling customers to build bespoke knowledge models.
In an interview during re:Invent 2025, Eshan Bhatnagar, Amazon’s Director of AGI Products, noted that while much of the industry attention remains fixated on performance benchmarks, AWS is increasingly emphasizing how Nova Forge can help enterprises construct deeply customized models that better reflect their operational realities.
Bhatnagar explained that AWS is committed to offering customers a comprehensive suite of AI resources. This is why the latest Amazon Bedrock expansion added eighteen widely used models and why AWS continues to update its proprietary Nova series according to evolving customer requirements.
Discussing Nova Forge specifically, Bhatnagar observed that organizations often struggle when implementing general-purpose models, as they must invest considerable time integrating proprietary knowledge and internal policies to ensure outputs align with real-world constraints. This process demands meticulous validation to prevent errors or unintended behaviors. Nova Forge addresses this by enabling an open-training paradigm in which customers may fine-tune the model at the pre-training, mid-training, or post-training stage, weaving domain-specific expertise directly into the model rather than treating it as an external attachment—an approach that frequently leads to inconsistent or unpredictable results.
He further emphasized that Nova Forge was engineered to make model customization far more flexible. As organizational rules and knowledge evolve, customers may adjust the model’s behavior through incremental fine-tuning at any stage, ultimately producing outputs that are more reliable and more closely aligned with operational expectations.
Complementing this, AWS continues to enhance its guardrails framework, allowing customers to define enforceable behavioral boundaries rather than relying solely on the model’s built-in heuristics.
Regarding model selection, Bhatnagar reiterated that AWS positions itself as a provider of choice rather than prescription. Customers may continue adopting the latest industry-standard models or rely on the evolving Nova family. Ultimately, the decision rests with the customer—AWS does not intend to dictate which models should be used.
Given Bhatnagar’s previous leadership of the Alexa team, the interviewer asked whether Nova Forge could be applied to current Alexa services, particularly since the new Alexa+ is natively powered by Nova models. Bhatnagar confirmed that developers can indeed tailor Alexa behaviors directly through Nova Forge without modifying the underlying Nova model, meaning Alexa for Business deployments may already benefit from targeted fine-tuning.
As for whether the Nova models’ native support for more than 200 languages—including Chinese—implies future Chinese-language interaction in Alexa+, Bhatnagar responded that such decisions will depend on market demand, and that there are currently no plans to introduce full Chinese-language support to Alexa+.
Donate