Amazon Has New Frontier AI Models—and a Way for Customers to Build Their Own

Steven

Amazon Has New Frontier AI Models—and a Way for Customers to Build Their Own

Summary

Amazon unveiled the second generation of its Nova family at re:Invent: Nova 2 Lite and Nova 2 Pro (large language models), Nova Sonic (real-time voice), and Nova Omni (a fully multimodal model that accepts images, audio, video and text and performs simulated reasoning). Alongside those models Amazon launched Nova Forge — a tool that lets customers add their own data at multiple stages of training (including custom pretraining) to produce specialised frontier models hosted on AWS.

Early testers include Reddit (built a Reddit-specific moderation model), Booking.com, Sony and Nimbus Therapeutics. Amazon claims Nova 2 Pro matches or exceeds leading rivals on several benchmarks and says Nova Forge should be far cheaper than building a frontier model from scratch, although precise costs were not disclosed. The approach trades flexibility and deeper customisation for tighter AWS cloud lock-in.

Key Points

  1. Nova 2 family: Nova 2 Lite and Nova 2 Pro are Amazon’s new LLMs; Nova Sonic is a real-time voice model; Nova Omni is a fully multimodal reasoning model.
  2. Nova Forge lets customers inject their own data during various training stages, including custom pretraining — not just conventional fine-tuning.
  3. Amazon positions Forge as a way to build domain-expert frontier models more cheaply than building from scratch, with AWS as the host.
  4. Reddit used Nova Forge to create a specialised model for content-moderation tasks that standard off-the-shelf models struggle with.
  5. Other early testers: Booking.com, Sony and Nimbus Therapeutics — showing cross-industry interest (travel, entertainment, biotech).
  6. Amazon claims Nova 2 Pro rivals GPT-5/GPT-5.1, Gemini Pro 2.5/3.0 Pro and Anthropic’s Sonnet 4.5 on benchmarks and highlights agentic/task-following strengths.
  7. Trade-offs: deeper customisation and potentially better domain performance, but increased dependence on AWS infrastructure and possible data/control considerations.

Context and Relevance

This launch sits at the intersection of two big trends: enterprises wanting AI models that truly understand their domain, and cloud providers racing to offer not just inference but full training/customisation pipelines. Many companies find off-the-shelf models insufficient for specialised workflows; Nova Forge’s ability to add data during pretraining is notable because that stage historically required vast resources and lab-grade engineering.

Amazon is playing catch-up in frontier LLMs but is leveraging AWS scale and customers. The move matters for organisations weighing the cost and benefits of custom models versus open-model experimentation — and for the competitive dynamics between AWS, Google Cloud, Microsoft/Azure, and rising players that bundle models with cloud services.

Why should I read this?

Quick and blunt: if you care about making AI actually useful for your business (not just pretty demos), this matters. Nova Forge isn’t just another fine-tuning tool — it lets firms push domain data deeper into the model-building process, so you can get a proper expert model without paying the cost of training from scratch. It’s AWS-centric, so expect wins for capability but think about lock-in and data governance first. Read it if you’re deciding how to build or buy bespoke AI.

Source

Source: https://www.wired.com/story/amazon-nova-forge-ai-models/