Amazon is trying to make its AI pitch less about one model for everyone and more about models shaped for a company’s own work. At re:Invent in Las Vegas, the company announced a second generation of Nova AI models alongside Nova Forge, a tool that lets customers add their own training data to unfinished Nova models.
The strategy matters because Amazon’s Nova models are not as widely used as models from OpenAI and Google. But Amazon has a different advantage: its cloud business. If customers can build more specialized frontier AI models inside Amazon’s cloud, Nova could become more relevant to companies that already rely on Amazon infrastructure.
A new Nova lineup
Amazon detailed several additions to the Nova family. Nova Lite and Nova Pro are improved large language models. Nova Sonic is a real-time voice model. Nova Omni is a more experimental model that uses images, audio, and video as well as text to perform a simulated form of reasoning.
The new models are being made available today to a limited number of customers. Amazon says Nova 2 Pro matches or exceeds OpenAI’s GPT-5 and GPT-5.1, Google’s Gemini Pro 2.5 and Gemini 3.0 Pro, and Sonnet 4.5 from Anthropic across a range of benchmarks. Rohit Prasad, who leads Amazon’s AI efforts, says Nova 2 Pro is especially strong at agentic tasks, including following complex instructions and using tools on a computer.
Amazon also says Nova 2 Lite is similar to Claude 4.5 Haiku, GPT-5 Mini, and Gemini Flash 2.5 on various benchmarks. Those comparisons are important to Amazon’s positioning, but the larger strategic point is customization. The company is not only introducing models; it is offering a route for customers to help shape them.
What Nova Forge changes
Nova Forge is designed to let customers build specialized frontier models by adding their own training data to unfinished versions of Nova 2 Lite and Pro. That is different from ordinary fine-tuning, where a company adjusts a completed model after the main training process is already finished.
With Nova Forge, customers can add data at multiple points in model training. The most notable stage is custom pretraining, which involves the base model itself. That kind of work is usually associated with large AI labs, not ordinary enterprise AI teams.
Prasad told WIRED, “Everyone is looking for a frontier model that's an expert in their domain.” He said Amazon first developed the technologies behind Nova Forge for internal teams, including those working on Alexa and AI agents. He described the approach as “essentially a new open training paradigm.”
The practical appeal is straightforward. Many companies do not just want a capable general-purpose AI system. They want a model that understands their specific data, workflows, policies, and edge cases well enough to be useful in production.
Reddit shows the use case
Reddit has already tested Nova Forge. The company used it to build a custom model for identifying content that violates platform rules. According to Reddit chief technology officer Chris Slowe, a standard fine-tuning approach would not have been enough for that task.
Slowe said most models are built to avoid offensive or violent content entirely, which can make them refuse to analyze some material. For a content platform, that limitation matters. A moderation model may need to understand difficult material in order to evaluate it against rules.
By combining custom pre-training with conventional fine-tuning, Reddit created what Slowe called a model with deeper knowledge of the platform. He told WIRED, “Other LLMs understand Reddit as a concept, and how Reddit works, but they're not down in the weeds.” He added, “We really built a Reddit expert model.”
Slowe said the customized model could support a range of uses, and will most likely next be used to automate content moderation. Other companies testing Nova Forge include Booking.com, Sony, and Nimbus Therapeutics, a biotech firm.
Why cloud customers may care
Amazon’s approach arrives as companies are looking for AI tools that go beyond the latest general-purpose models. A survey from the consulting company Bain released in November found that about three-quarters of US companies see AI as a high priority. At the same time, those companies report problems using AI, including a lack of expertise and resources needed to build custom models.
That gap is where Nova Forge could fit. Building a large language model from scratch can cost tens or hundreds of millions of dollars. Prasad says a frontier model built with Nova Forge should be significantly cheaper, though he did not give specific figures.
The tradeoff is that Nova Forge is tied to Amazon’s cloud. Today, companies often choose between closed models, which are accessed through an API or app, and open models, which can be downloaded and run on a company’s own hardware. Open models can be cheaper to experiment with and easier to modify, but the data used to train them is typically not released, which makes tuning more limited and complicated.
Nova Forge offers another path: more access to the training process, but within Amazon’s cloud environment. For Amazon, that could make AI customization part of the broader competition for cloud customers.
Amazon’s broader AI bet
Amazon remains a darker horse in the AI race because it was relatively late to develop truly cutting-edge AI language models. Still, the company is building a wider set of advanced AI capabilities. It has added generative AI to its shopping platform, including the ecommerce-focused chatbot helper Rufus.
Like other major technology companies, Amazon is investing billions in AI infrastructure, betting that demand for AI will keep growing quickly. It is competing with Google and Microsoft for cloud customers. OpenAI is also building infrastructure rapidly and could someday become a cloud player itself.
Amazon has also invested $8 billion in Anthropic, a major OpenAI competitor founded by staff who left the maker of ChatGPT. At the hardware level, Amazon is trying to challenge Nvidia’s dominance, and Anthropic’s latest models are trained on Amazon’s custom Trainium chips.
Nova 2 Omni points to another part of Amazon’s ambition. The model can take images, audio, video, and text as input and use simulated reasoning to generate output. Prasad says that, to his knowledge, no other AI company has released a fully multimodal model of this kind.
The strongest message from Amazon’s announcement is not only that it has new AI models. It is that the company wants the future of enterprise AI to be more specialized, more closely tied to customer data, and more deeply connected to cloud infrastructure.