Nova Forge lets Amazon customers train frontier AI for tasks, promising practical gains for businesses, but there’s a surprising twist…

Nova Forge gives Amazon customers a way to train frontier models for specific tasks, a change that could make AI more practical for many businesses.

Amazon rolled out a second generation of its Nova models at re:Invent in Las Vegas and introduced a set of related tools aimed at letting companies build their own frontier models. The new Nova family is not nearly as popular as rival offerings from OpenAI and Google, but the emphasis on customer customization could help the models gain traction among Amazon Web Services users.

The company described two improved large language models, Nova 2 Lite and Nova 2 Pro; a realtime voice model called Nova Sonic; and a more experimental option, Nova Omni, which applies a simulated kind of reasoning to images, audio, video and text. These models are being made available today to a limited number of customers.

Given the strategic importance of its cloud business, Amazon is also releasing a tool called Nova Forge that lets customers create specialized frontier models by adding their own training data into unfinished versions of the Nova 2 Lite and Nova 2 Pro models. Fine-tuning off-the-shelf models such as Google’s Gemini and OpenAI’s GPT has been possible for some time, yet Nova Forge lets teams inject data at multiple stages of training, including during base-model construction in a process known as custom pre-training — something that typically happens inside large AI labs.

“Everyone is looking for a frontier model that's an expert in their domain,” Rohit Prasad, who leads Amazon’s AI efforts, said ahead of the announcements. He said Amazon developed the underlying Forge technologies to give internal teams the same capabilities they were offering customers, naming groups that work on Alexa and AI agents as examples. He described the setup as "a new open training paradigm."

Reddit is one customer that has tried the approach. The company used Nova Forge to build a custom model designed to identify content that violates platform rules. Chris Slowe, Reddit’s chief technology officer, said fine-tuning a conventional model would not work since most models are built to avoid offensive or violent content entirely and would refuse to analyze some materials. Custom pre-training, combined with standard fine-tuning, produced a frontier model tuned to Reddit’s particular style and content.

“Other LLMs understand Reddit as a concept, and how Reddit works, but they're not down in the weeds,” Slowe says. “We really built a Reddit expert model.” He expects the customized model will have multiple applications and will likely be deployed alongside existing systems to help automate content moderation.

Booking.com, Sony and Nimbus Therapeutics, a biotech company, are among the other organizations testing Nova Forge.

That ability to craft specialized models may appeal to firms that need more than broad, general-purpose systems. A November survey from Bain found roughly three-quarters of U.S. companies treat AI as a high priority. Those organizations report a range of obstacles to adoption, including limited expertise and the resources required to build and maintain custom models.

Most production AI models fall into one of two camps: closed models that are accessed via an API or an app, or open models that can be downloaded and run on a company’s own hardware. Many firms pick open models — popular examples include offerings from Chinese companies like Alibaba and DeepSeek — because they are cheaper to experiment with and easier to modify. The training data for many open models is not released, though, which makes targeted tuning harder.

Nova Forge represents a different approach, locked into Amazon’s cloud. Building a large language model from scratch can cost tens of millions or more; Prasad said a frontier model assembled with Nova Forge should be far less expensive, though he declined to provide a price estimate.

Amazon has often been seen as a late entrant to cutting-edge AI language models, yet it has been quietly expanding a set of advanced capabilities across its businesses. The company has added generative AI into its e-commerce products, for example in a shopping assistant called Rufus. It is also investing heavily in data centers and custom hardware as part of a major bet that demand for AI will keep growing.

The company faces stiff competition for cloud customers from Google and Microsoft. OpenAI is rapidly building its own infrastructure and could enter the cloud market itself; it has hedged its bets by investing $8 billion in Anthropic, a major competitor to OpenAI founded by staff who left the maker of ChatGPT. Amazon is also looking to challenge Nvidia’s hold on training hardware, and Anthropic’s latest models have been trained on Amazon’s custom Trainium chips.

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. Prasad pointed out that the model is especially capable at agentic tasks such as following complex instructions and using tools on a computer. The company claims Nova 2 Lite performs similarly to Claude 4.5 Haiku, GPT-5 Mini and Gemini Flash 2.5 on several benchmark tests.

Nova Omni is a more experimental offering that makes a clear case Amazon is active in core AI research. The model is fully multimodal: it accepts images, audio, video and text as input and runs a form of simulated reasoning to produce outputs. Prasad said, to his knowledge, no other AI company has released a fully multimodal model of this kind.

For Reddit’s Chris Slowe, the most valuable feature of Nova is how customizable it can be. “I do believe it has a lot of potential,” he says. “For a large set of situations, it will be substantially better than what we get off-the-shelf.”