Imagine having the power to shape AI to perfectly fit your needs. Amazon is making waves in the AI world, not just by creating new models, but by handing the reins over to you! They've unveiled a new family of AI models, and more excitingly, a way for customers to essentially build their own, highly specialized AI powerhouses.
While Amazon's AI models haven't yet reached the household name status of OpenAI's or Google's offerings, their latest move could be a game-changer, especially for their cloud computing clients. At re:Invent, their Las Vegas conference, Amazon announced the second generation of its Nova AI models.
Let's break down what's new: two improved large language models called Nova Lite and Nova Pro, a real-time voice model named Nova Sonic, and a fascinating experimental model, Nova Omni. Nova Omni is designed to simulate reasoning using a combination of images, audio, video, and text. These models are rolling out to a select group of customers right now.
But here's where it gets really interesting... Amazon is launching Nova Forge. And this is the part most people miss. It's a tool that allows customers to create highly customized AI models. How? By adding their own specific training data to unfinished versions of the Nova 2 Lite and Pro models. Think of it as getting a partially built engine and then being able to add your own custom parts to make it run exactly the way you want.
Now, fine-tuning existing AI models like Google's Gemini or OpenAI's GPT is already possible. But Amazon's approach goes much deeper. It allows customers to inject their data at various stages of the model's development, including the crucial process of building the base model – something known as custom pretraining. This level of control was traditionally reserved for the elite AI labs.
According to Rohit Prasad, who spearheads Amazon’s AI initiatives, the technology behind Nova Forge was initially developed to empower internal teams, including those working on Alexa and AI agents. He describes it as “essentially a new open training paradigm.”
One early adopter is Reddit. They used Nova Forge to create a custom model specifically designed to identify content that violates the platform's rules. Standard fine-tuning wouldn't cut it, explains Reddit's chief technology officer, Chris Slowe. Why? Because most AI models are trained to avoid offensive or violent content altogether, which means they'd refuse to even analyze the type of content Reddit needed to moderate. Custom pre-training, combined with traditional fine-tuning, allowed Reddit to build an AI model that's truly an expert in understanding and navigating the nuances of their platform.
“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 envisions a range of uses for their customized model, with automated content moderation being the most likely next step. Other companies currently testing Nova Forge include Booking.com, Sony, and Nimbus Therapeutics, a biotech company.
This ability to create highly specialized models could be a smart move by Amazon. Many companies are seeking AI tools that surpass the limitations of general-purpose models. A Bain & Company survey revealed that about three-quarters of US companies consider AI a high priority. However, these same companies also report significant challenges in using AI, including a lack of expertise and resources to build custom models.
Today, AI models generally fall into two categories: closed (accessible only through APIs or apps) and open (downloadable and runnable on your own hardware). Many companies prefer open models, particularly those from Chinese companies like Alibaba and DeepSeek, because they're more affordable for experimentation and easier to modify. However, the training data for these open models is usually not released, which limits the potential for fine-tuning.
And this is a controversial point: Nova Forge offers a different path, but it's one that's firmly rooted within Amazon's cloud infrastructure. Building a large language model from scratch can cost tens, even hundreds, of millions of dollars. Prasad claims that using Nova Forge will significantly reduce these costs, though he didn't provide specific figures.
Amazon might be considered a latecomer to the AI race, but they're steadily building a powerful portfolio of AI capabilities. They've also integrated generative AI into their shopping platform with tools like Rufus, an e-commerce-focused chatbot.
Like its competitors, Amazon is pouring billions into building new AI infrastructure, a massive bet on the continued growth of AI demand. They're competing fiercely with Google and Microsoft for cloud customers. And, let's not forget OpenAI, which is rapidly building its own infrastructure and could one day become a cloud player itself. Amazon has hedged its own bets by investing heavily in Anthropic, a major OpenAI competitor. They're also challenging Nvidia's dominance in hardware, with Anthropic's latest models being trained on Amazon's custom Trainium chips.
Amazon claims that Nova 2 Pro either matches or exceeds the performance of OpenAI’s GPT-5 and GPT-5.1, Google’s Gemini Pro 2.5 and Gemini 3.0 Pro, and Anthropic's Sonnet 4.5 across various benchmarks. Prasad emphasizes its strength in agentic tasks, such as following complex instructions and using computer tools. They also state that Nova 2 Lite is comparable to Claude 4.5 Haiku, GPT-5 Mini, and Gemini Flash 2.5 on a variety of benchmarks.
Nova 2 Omni demonstrates Amazon's growing strength in AI research. This multimodal reasoning model can process images, audio, video, and text, and perform simulated reasoning to generate output. Prasad believes that no other AI company has released a fully multimodal model of this type.
According to Reddit's Slowe, the customizable nature of Nova is its most valuable asset. “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.”
So, what do you think? Is Amazon's approach of empowering customers to build their own AI models a game-changer, or is it too restrictive to be truly revolutionary? Will this put pressure on other major players to offer similar customization options? And, more importantly, do you believe that custom pre-training is a necessity for businesses looking to truly leverage the power of AI, or is fine-tuning enough? Share your thoughts in the comments below!
