AWS Debuts Nova 2 AI Models and Nova Forge

Amazon Web Services has announced a significant expansion of its homegrown AI portfolio, including the Nova 2 collection of foundational models and the Nova Forge to allow enterprises to build custom “frontier” variants with more control over data and behavior. The twin launch drives AWS deeper into agentive AI and multimodal systems while addressing long-standing enterprise requirements around personalization, governance, and pricing predictability.

What’s new in Nova 2: models for multimodal reasoning

For that, the Nova 2 series comes in as four separate models, each meant to cater to a range of requirements. Nova 2 Lite focuses on simple automation with a low-cost reasoning model that can interpret text, images, and video. Nova 2 Pro takes that a step further with agentive design, processing even more complex workflows that involve code generation and tool use across text, images, video, and speech.

Nova 2 Sonic is geared for voice-first, low-latency experiences to support speech-to-speech conversations. Finally, completing the set is a fully multimodal model called Nova 2 Omni that can take text, images, video, and speech in and out, producing both text and images—perfect for advanced copilots, media workflows, and analytics.

Together, the models represent the direction that enterprise AI is moving: reasoning-centric systems that can plan, call tools to aid it, and fuse multimodal inputs together instead of just completing text. That transition opens up higher-value use cases—coding bots that read logs and screenshots, commerce agents that understand product photos, service bots that deal with voice, forms, and documentation.

Nova Forge lets customers create custom ‘Novella’ models

Nova Forge is the wider strategic swing. Per CNBC, businesses can subscribe for $100,000 a year to iterate on Nova and create their own variations of it—referred to as Novellas—starting from pre-trained, mid-trained, or post-trained points. And that flexibility is important: many companies would like to go further than simple retrieval or light fine-tuning and bake in domain expertise earlier in the training pipeline—not just tack it on afterward.

This approach also addresses a well-known issue in the literature: catastrophic forgetting. As new domain data is ingested by models in further stages of training, such models may gradually forget core reasoning and general knowledge. Nova Forge allows for controlled access to various points in the training arc, and is built to maintain baseline competencies while safely introducing proprietary datasets, policies, and toolchains.

Why enterprise control and multimodal reasoning matter

Enterprises discovered that retrieval-augmented generation alone is insufficient to meet accuracy, fairness, or latency requirements at scale. They need models that can think between modalities, invoke internal tools, and respect business rules while keeping data residency and audit needs in check. Nova 2’s agentic focus and Nova Forge’s more granular customization hope to reach that bar without offloading end-to-end model training complexity in the lap of companies.

The timing coincides with an overall market trend. Gartner has predicted swift adoption of generative AI by businesses, at least through the middle of the decade, while McKinsey estimated that it could annually juice economic output between $2.6 trillion and $4.4 trillion worldwide. As budgets tighten, CIOs are focusing on observability, cost control, and leveraging of vendors—categories in which a bespoke cloud-native model stack can differentiate.

Competitive positioning against OpenAI, Anthropic, Google

AWS has a two-front battle: on model quality, and deployment control. Rivals like OpenAI, Anthropic, and Google are providing ever more powerful reasoning and multimodal models—Yeung is friends with the founders of all three companies—as well as special training courses for their top customers. AWS’s pitch is to connect home-grown models with its cloud primitives—storage, data lakes, security, MLOps/machine learning operations, and scale economics—so that customers can pull in data from services like S3 and Redshift without giving up governance into a single pipeline.

Scale remains a strategic asset. AWS continued to account for about a third of the global spend in cloud infrastructure, according to industry monitoring org Synergy Research—and that is its runway, offering the firm distribution, integration reach, and cost levers that ain’t there for pure-play model vendors. If Nova Forge is able to make such custom training smooth on that footprint, it stands to be a well-trodden path for large enterprises that are modernizing analytics and software delivery with agentic AI.

Early users and use cases across media, travel, and tech

Early Nova Forge interest came from Reddit, Sony, and Booking.com, according to AWS—brands that have varied content and service requirements, including for content operations, media localization, trust and safety needs, and tailored travel search. Look for Nova 2 Sonic to appear in contact centers and voice assistants, Nova 2 Pro in developer platforms and IT automation, and Nova 2 Omni in knowledge management as well as marketing creative pipelines that combine text, imagery, and video.

AWS also has claimed that it already serves tens of thousands of customers with the Nova family, so the company appears likely to seed Nova 2 upgrades via its installed base rather than start from scratch.

For businesses, that opportunity translates to relatively less friction in piloting multimodal and agentic capabilities alongside familiar security and monitoring patterns.

What to watch next: benchmarks, latency, and workflows

Key landmarks will be independent benchmarks for reasoning and tool use; latency for speech-to-speech interactions; and operational workflow for creating Novellas—dataset governance, evaluation frameworks, rollback safety, and cost transparency. How closely Nova 2 and Nova Forge integrate with common data sources and AIOps stacks will impact how quickly these products can get out of pilot mode and into production.

The headline feels obvious: AWS is wagering that multimodal reasoning plus customer-controlled customization will finally be the recipe for enterprise AI success. If Nova 2′s features survive the rigors of the real world, and Nova Forge can facilitate high-quality custom training without runaway costs, it could enable the company to substantially expand its AI footprint inside the world’s largest IT budgets.