Introduction
Amazon Web Services (AWS) has long been a dominant force in cloud infrastructure, but its latest announcement signals a deeper commitment to the next wave of artificial intelligence: custom, foundation‑class models that can be tailored to an organization’s unique data without the need for expensive GPU clusters. The service, dubbed Nova Forge, sits atop AWS’s newly released Nova 2 family of models and promises a streamlined path from proprietary data ingestion to a production‑ready model that can be deployed on Bedrock, AWS’s managed AI platform. For businesses that have struggled to keep pace with the rapid evolution of large language models (LLMs) due to hardware constraints, Nova Forge offers a compelling alternative. By integrating proprietary data at every stage of training—what AWS calls “open training”—the service mitigates the risk of catastrophic forgetting while preserving the foundational capabilities of the underlying Nova models. This approach addresses a key pain point for enterprises: the ability to embed deep domain knowledge into a model that still retains general reasoning, instruction following, and code execution skills.
The announcement came alongside the rollout of Nova 2 Lite, Nova 2 Pro, Nova 2 Sonic, and Nova 2 Omni, each designed to tackle different modalities and workloads. Nova 2 Lite, for instance, is a cost‑effective reasoning model that can process text, images, and video to generate text, while Nova 2 Pro offers advanced reasoning for coding, long‑range planning, and problem solving. Nova 2 Sonic adds speech‑to‑speech capabilities, and Nova 2 Omni brings multimodal generation to the table. Together, these models provide the foundation upon which Nova Forge builds custom “Novellas,” the proprietary versions of Nova that enterprises can train and deploy.
In the sections that follow, we’ll unpack how Nova Forge works, the technical advantages it offers, real‑world use cases, and what this means for the future of enterprise AI.
Custom Model Creation with Nova Forge
At its core, Nova Forge is a managed service that removes the operational overhead of building and fine‑tuning large language models. Traditionally, organizations that wanted to create a foundation model had to invest in high‑performance GPUs—often multiple Nvidia H100s—alongside a team of data scientists and machine learning engineers. Nova Forge eliminates that barrier by allowing developers to blend proprietary data with an Amazon‑curated dataset at every step of the training pipeline. The service provides checkpoints, so teams can monitor progress and ensure that the model does not regress on its baseline capabilities.
The concept of “open training” is particularly noteworthy. Instead of a single fine‑tuning pass, Nova Forge encourages continuous integration of new data, enabling the model to learn domain‑specific nuances while still retaining the general knowledge embedded in the original Nova architecture. This iterative approach is analogous to how humans learn: we constantly absorb new information without losing what we already know. By embedding proprietary data early in the training process, enterprises can produce a model that feels native to their business context—whether that’s a legal firm that needs to understand regulatory language or a financial institution that must interpret complex market data.
Once a Novella is trained, it can be imported into Bedrock, where it becomes available as a managed endpoint. Bedrock’s API layer abstracts away the complexities of scaling, monitoring, and securing the model, allowing developers to focus on building applications and agents that leverage the newly minted AI capabilities. For example, a retail chain could deploy a Novella that understands its product catalog, pricing strategy, and customer service scripts, and then expose it through Bedrock to power a conversational agent that assists shoppers in real time.
The Nova 2 Family: Powering the Forge
The Nova 2 family serves as the backbone for Nova Forge, and each variant brings unique strengths to the table. Nova 2 Lite is positioned as a workhorse for everyday tasks, offering a balance between performance and cost. Benchmark results show that Nova 2 Lite matches or surpasses competitors such as Claude Haiku 4.5, GPT‑5 Mini, and Gemini Flash 2.5 on a majority of tests, making it an attractive option for production workloads.
Nova 2 Pro, on the other hand, is the most capable reasoning model in the lineup. It excels at multi‑document analysis, video reasoning, advanced mathematics, and agentic engineering tasks. Its built‑in grounding and code execution capabilities mean that it can not only generate text but also run code snippets in real time—a feature that is increasingly valuable for developers building AI‑powered tools.
Nova 2 Sonic expands the modality spectrum to include speech‑to‑speech, supporting multiple languages and a 1‑million‑token context window. This allows for sophisticated voice interactions that can switch topics mid‑conversation, a feature that could be leveraged in customer support or virtual assistant scenarios.
Finally, Nova 2 Omni is the most ambitious model, capable of simultaneously analyzing text, images, video, and audio. With a context window that can handle up to 750,000 words or hours of audio, Omni can ingest entire product catalogs, brand guidelines, and video libraries in a single pass. This multimodal prowess positions Omni as a powerful tool for industries that rely heavily on visual content, such as media, advertising, and e‑commerce.
The synergy between these models and Nova Forge is evident: enterprises can choose the variant that best aligns with their use case, train it with proprietary data, and then deploy it on Bedrock. The result is a highly specialized AI that still retains the robustness of a foundation model.
Real‑World Impact and Use Cases
Reddit is one of the first companies to showcase the practical benefits of Nova Forge. By integrating its own community data and moderation guidelines into a custom Novella, Reddit has built a moderation program that can understand the nuances of its user base and enforce policies more effectively. This example illustrates how domain knowledge—often the most valuable asset for a company—can be embedded into an AI model without compromising its general reasoning abilities.
Other potential use cases span a wide spectrum. In healthcare, a hospital could train a Novella on patient records, clinical guidelines, and medical literature to assist clinicians with diagnosis and treatment recommendations. In finance, a bank could embed regulatory compliance data to ensure that automated advisory services stay within legal boundaries. In manufacturing, a Novella could be trained on equipment specifications and maintenance logs to power predictive maintenance agents.
Because Nova Forge only works with Nova models, AWS is currently limiting the service to its own ecosystem. However, the company has indicated plans to expand support to other Nova 2 variants beyond Lite. This expansion will broaden the range of capabilities available to enterprises, from advanced reasoning to multimodal generation.
Technical Considerations and Future Directions
While Nova Forge removes the need for GPU clusters, organizations must still consider data governance, privacy, and security. The proprietary data that feeds into the training pipeline must be handled carefully to avoid accidental exposure. AWS’s Bedrock platform offers built‑in compliance features, but companies should audit their data pipelines to ensure alignment with regulations such as GDPR or HIPAA.
Another consideration is the trade‑off between customization and generality. The more domain‑specific a Novella becomes, the more it may diverge from the baseline Nova model’s behavior. AWS mitigates this risk through checkpointing and continuous monitoring, but developers must still validate that the model performs as expected across a range of scenarios.
Looking ahead, the concept of “reinforcement learning gyms” introduced by Nova Forge hints at a future where enterprises can simulate complex environments to fine‑tune agents in a controlled setting. This could accelerate the development of specialized AI assistants that can navigate intricate workflows, such as legal document review or supply chain optimization.
Conclusion
AWS’s Nova Forge represents a significant shift in how enterprises can approach AI model development. By marrying the power of foundation models with a managed, GPU‑free training pipeline, the service democratizes access to advanced AI capabilities. The ability to embed proprietary data at every step of training, coupled with the flexibility to deploy on Bedrock, offers a compelling proposition for businesses that need to stay competitive in an AI‑driven world.
The Nova 2 family’s diverse modalities—from text and image to speech and multimodal reasoning—mean that nearly every industry can find a suitable base model to customize. As more enterprises adopt Nova Forge, we can expect to see a wave of AI applications that are both deeply knowledgeable about their domain and robust enough to handle general reasoning tasks.
Call to Action
If your organization is looking to build a custom AI model without the overhead of GPU infrastructure, now is the time to explore AWS Nova Forge. Start by evaluating which Nova 2 variant aligns with your business needs, then experiment with a small dataset to gauge the model’s performance. Reach out to AWS’s Bedrock team for guidance on data preparation and deployment best practices. By leveraging Nova Forge, you can unlock the full potential of foundation‑class AI while keeping costs and complexity under control.