Key Takeaways

• AWS has launched Nova 2, an advanced suite of AI models focused on reasoning, multimodal input, and real-time voice applications, catering to diverse needs from budget-friendly options to sophisticated capabilities.
• Alongside Nova 2, AWS revealed Nova Forge, a platform that enables users to create custom Nova variants—termed “Novellas”—by integrating proprietary data earlier in the training process to mitigate issues like catastrophic forgetting.
• Major brands are already engaging with these models, showcasing a strong demand for enhanced control over AI behavior and knowledge.

Why This Launch Matters

For those who feel enterprise AI has been confined between overly generic APIs and costly, custom-built models, AWS’s new products aim to establish a balanced middle ground. Nova 2 revitalizes Amazon’s AI model lineup, enhancing its reasoning and multimodal capabilities, while Nova Forge allows companies to incorporate their unique expertise early in the training without incurring the full costs of traditional model development. This represents a significant step towards greater control over AI systems, particularly for teams prioritizing accuracy, safety, and long-term maintainability.

---

What AWS Announced at re:Invent 2025

On December 2, 2025, during CEO Matt Garman’s keynote at re:Invent, AWS unveiled four distinct models under the Nova 2 umbrella. These new models aim to support tasks ranging from everyday assistance to complex problem-solving, as well as real-time communication and rich multimodal reasoning.

Overview of the Nova 2 Lineup

• Nova 2 Lite: A swift and budget-conscious reasoning model equipped with developer tools for customizable cognitive capabilities. It accommodates tools like code interpreters and web grounding, alongside a 1M-token context window for extended tasks. Common applications include customer service chatbots and task automation.
• Nova 2 Pro (Preview): A powerful model designed for more complex reasoning and agent tasks, especially code-related duties. Early access is available to Nova Forge clients.
• Nova 2 Sonic: A speech-to-speech model facilitating natural, real-time conversational AI with features like polyglot voices and cross-modal interactions, featuring a 1M-token context window across seven languages and seamless integration with telephony.
• Nova 2 Omni (Preview): A multimodal reasoning and generation model that can process text, images, video, and audio, generating coherent outputs across these formats.

AWS positions Nova 2 as a versatile toolkit for agent-based AI that transcends chat, enabling enterprises to transition from simple demos to robust production systems capable of handling documents, videos, user interactions, and executing actions reliably.

The Promise of Enhanced Reasoning

Designed for planning, decomposition, and resolution of multi-step tasks, Nova 2’s reasoning models go beyond mere word prediction. They combine long context windows with tool utilization and safety features. For instance, an AI agent handling claims processing might read PDFs, verify policy tables, engage internal tools, and produce verifiable summaries within a single workflow.

---

Nova Forge: Crafting Your Own “Novellas”

The other significant element of this launch, Nova Forge, allows users to fine-tune models more comprehensively. Instead of merely making adjustments to a fully-trained model, Nova Forge allows for the integration of proprietary data during earlier training stages—be it pre-, mid-, or post-training—thus avoiding the common issue of models forgetting prior acquired knowledge. AWS refers to this innovative method as an “open training” approach, which retains broad capabilities while integrating specific domain knowledge effectively.

How It Works in Practice:

1. Choose Your Starting Point: Select from early training checkpoints based on your specific needs for domain knowledge infusion.
2. Train Within Your Environment: Run reinforcement fine-tuning while employing your own reward functions and implement safety protocols.
3. Utilize Familiar AWS Services: Build using SageMaker AI and deploy via Bedrock, with early access to models such as Nova 2 Pro and Nova 2 Omni for Forge customers.

Enterprises operating in regulated or niche industries often require models familiar with specific jargon and workflows, all while ensuring a broad understanding. By customizing earlier, Novellas strive to blend general knowledge with the intricacies relevant to your organization.

Pricing and Availability

AWS has indicated that Nova Forge operates on an annual subscription basis, although specific pricing details are only accessible through the console. While some reports have mentioned a potential annual fee, the official AWS pricing page does not currently disclose a specific amount. Nova Forge has initially launched in the US East (N. Virginia) region, with more geographical availability planned.

Who’s Getting Started

Early adopters include noted companies like Reddit, Sony, and Booking.com, reflecting a blend of consumer and enterprise-level applications. Additionally, there are pilot programs underway in the healthcare and biotech sectors. Should these initiatives prove successful, a noteworthy trend may emerge: a shift from numerous task-specific models to a more streamlined array of high-reliability agents built on Novellas.

---

How Nova 2 Fits Into the Broader AI Ecosystem

For organizations evaluating performance, control, and cost, Nova 2 and Nova Forge offer a balanced alternative:

• Closed, Hosted Models: Very convenient but often challenging to customize deeply or to manage data in a compliance-centric manner.
• Full Custom Training: Provides the ultimate level of control but demands a significant investment of time and resources.

AWS aims to deliver robust reasoning and multimodal capabilities via Bedrock while enabling earlier data integration for training. This also leverages AWS’s infrastructure advancements (such as Trainium) for optimized throughput and energy efficiency.

Day-to-Day Control Features

• Early Data Integration: Helps prevent catastrophic forgetting when incorporating new knowledge.
• Enhanced Oversight: Facilitates the implementation of custom safety rules and reward functions aligned with your organization’s objectives.
• Diverse Model and Tool Selection: Allows users to blend Nova with Bedrock’s multi-model ecosystem, enabling component swaps without extensive app rewrites.

---

Practical Use Cases for Nova 2 and Nova Forge

Here are a few example scenarios that take advantage of Nova 2’s capabilities alongside Nova Forge’s customization potential:

Enhanced Customer Support with Contextual Awareness

1. Intake: Nova 2 Sonic provides real-time voice triage, accurately identifying customer intent across various languages.
2. Understanding: Leverage Nova 2 Lite or Pro to process extensive historical data and documents with a 1M-token context window.
3. Action: The AI agent utilizes approved tools (e.g., CRM, billing systems) to write clear and auditable resolutions.
4. Upgrade Path: Train Novellas specific to your terminology and compliance rules, allowing the agent to learn from experience instead of starting anew each time.

Software Engineering Assistants Familiar with Your Stack

1. Reasoning: Utilize Nova 2 Pro (preview) for planning multi-step tasks and code-focused reasoning.
2. Enterprise Fit: Develop a Novella trained on your repositories, design documents, and style guidelines while implementing safety and licensing controls through Forge.

Knowledge-Intensive Workflows

1. Multimodal Input: Nova 2 Omni (preview) can analyze manuals, screenshots, and videos to generate text or images for instructions and checklists.
2. Long Context Management: Handle large contracts or extensive reports without the hassle of breaking them into smaller sections.

---

Governance, Safety, and Reliability

AWS frames Nova 2 and Nova Forge as a means to align AI development with enterprise governance standards. The following three elements stand out:

1. Early Customization: By promoting early-stage adjustments, users can integrate guardrails and reward functions reflective of their organizational policies, rather than retrofitting them later.
2. Integrated Observability: By utilizing SageMaker AI and Bedrock for deployment, teams benefit from baked-in monitoring, access controls, and data privacy features.
3. Enhanced Agent Reliability: While not detailed in the initial announcement, AWS’s ongoing updates hint at improved reliability for user interface-driven agents, suggesting potential pathways for Novellas.

---

Factors to Consider Before Adoption

• Define Target Outcomes: Pinpoint one or two high-impact workflows (e.g., claims processing, code evaluations) where enhanced reasoning and long context can yield significant value.
• Assess Data Readiness: Examine the proprietary datasets you plan to incorporate through Forge and decide their integration at various training stages.
• Establish Safety Measures: Clearly outline necessary reward functions and guardrails early in the process. Conduct red-team simulations to gauge readiness.
• Cost Analysis: Estimate the total cost of ownership considering training, deployment, and inference costs, including potential savings from shifting workloads to Nova 2 Lite or Sonic where appropriate.
• MLOps Integration: Clearly map the transition from SageMaker AI development to deployment on Bedrock, ensuring a CI/CD pipeline and observability are in place.

---

Frequently Asked Questions

What Are Novellas?

Novellas refer to tailored versions of Amazon’s Nova models created through Nova Forge. They enable the integration of your proprietary data throughout earlier training stages, preserving general reasoning while enhancing domain-specific understanding.

How is Nova Forge Different from Fine-Tuning?

Fine-tuning modifies a pre-existing model at the end of its training, which can lead to the loss of prior knowledge. Nova Forge facilitates the incorporation of data at various training checkpoints and allows for reinforcement fine-tuning with customized reward functions, enhancing capabilities effectively.

Which Nova 2 Models Are Currently Available?

As of now, Nova 2 Lite and Nova 2 Sonic are generally accessible, while Nova 2 Pro and Nova 2 Omni are in preview mode with early access granted through Nova Forge. Always verify model availability in your AWS console as these statuses can periodically change.

Does AWS Provide Pricing Information for Nova Forge?

AWS confirms that Nova Forge operates on an annual subscription model, though specific pricing details are only available through the console. Some sources have mentioned potential fees, but there is no definitive number listed publicly.

Where Can You Access Nova Forge?

Currently, Nova Forge is available in the US East (N. Virginia) region, with additional regions set to launch in the future. Check AWS documentation for updates on regional availability.

---

Bottom Line

With the launch of Nova 2 and Nova Forge, AWS directly addresses the increasing enterprise demand for comprehensive reasoning and multimodal capabilities while integrating specific corporate knowledge and rules as early as possible. If your organization struggles to mold generic models into domain experts, this approach promises to streamline processes and improve reliability—especially for agents expected to read, watch, listen, and act responsively. The coming quarters will reveal how effectively Novellas operate in real-world applications, but the clear trajectory shows a movement toward more control and less compromise.