By AmazonLearning Objectives
- Understand how Nova Forge enables custom frontier model training
- Identify when custom model training is valuable vs. using off-the-shelf models
- Compare Nova Forge to fine-tuning options on other cloud platforms
What Is Nova Forge?
Nova Forge is Amazon's platform for building custom frontier models from Nova checkpoints using your own proprietary data. Launched at re:Invent 2025, it lets enterprises mix their domain-specific data with Amazon's Nova foundation weights to create specialized models — a unique offering among cloud providers.
This sits above standard fine-tuning: rather than adjusting a few layers of an existing model, Nova Forge enables deeper customization that produces models specialized for your industry, terminology, and use cases.
💡Key Concept
Fine-tuning vs. custom model training: Fine-tuning adjusts an existing model's behavior using your data (like teaching a student new material). Nova Forge goes further — it creates a new model variant that deeply incorporates your proprietary knowledge into the model weights (like training a specialist from the ground up). The result is more deeply specialized but requires more data and compute.
✅Tip
Access Nova Forge: Available through Amazon Bedrock — aws.amazon.com/bedrock — requires an AWS account
Pricing & Access
| Component | Cost | Details |
|---|---|---|
| Nova Forge training | Usage-based | Pay for compute time during model training |
| Model hosting | Usage-based | Standard Bedrock inference pricing for custom models |
| Data preparation | Included | Tools for preparing training data within the SDK |
Core Capabilities
Custom Frontier Model Building
Start from Amazon Nova checkpoints (Nova 2 Lite, Nova 2 Pro) and mix your proprietary data to create specialized models. The resulting model understands your domain at a depth that generic models cannot match.
Nova 2 Family Foundation
The base models are strong: Nova 2 Lite (best price-performance for everyday tasks) and Nova 2 Pro (complex agentic tasks), both supporting 1 million token context. Nova Forge builds on these foundations.
Enterprise Data Mixing
Proprietary data mixing techniques let you control how much your data influences the final model vs. the original Nova weights — balancing domain specialization with general capability.
Bedrock Integration
Custom models trained via Nova Forge deploy directly to Amazon Bedrock with the same enterprise privacy guarantees, compliance certifications, and operational tools as standard Bedrock models.
Strengths
- Unique offering: No other cloud provider offers custom frontier model training from their own checkpoints
- Deep customization: Goes beyond fine-tuning to create genuinely specialized models
- Bedrock ecosystem: Custom models get same enterprise SLAs, compliance, and monitoring
- AWS compliance: FedRAMP High, HIPAA, SOC 2 Type II, ISO 27001 certifications apply
Limitations & Considerations
- AWS-only: Custom models are tied to the AWS/Bedrock ecosystem
- Data requirements: Custom model training requires substantial, high-quality proprietary data
- Cost: Custom frontier model training is significantly more expensive than fine-tuning
- Expertise needed: Requires ML engineering knowledge to prepare data and evaluate results
- Time to value: Training cycles are measured in days to weeks, not minutes
Key Takeaways
- Nova Forge enables custom frontier model training from Amazon Nova checkpoints — unique among cloud providers
- Goes beyond fine-tuning to create deeply specialized models using proprietary enterprise data
- Built on Nova 2 family (Lite + Pro, 1 million token context) with full Bedrock integration
- Best for enterprises with substantial proprietary data that need models specialized beyond what fine-tuning achieves
