Build optimized variants of Nova called 'Novellas' by blending proprietary data with Nova's frontier capabilities. Start from early checkpoints across pre-training, mid-training, and post-training phases for maximum customization control.

Execute Reinforcement Fine-Tuning with reward functions in your own environment. Train models with custom reward signals to optimize for specific behaviors, outcomes, or quality metrics unique to your use case.

Implement custom safety guardrails using Nova Forge's built-in responsible AI toolkit. Define and enforce safety policies, content moderation rules, and alignment constraints specific to your deployment context.

Import custom Nova models as private models on Amazon Bedrock. Leverage the same security, consistent APIs, and broader AWS integrations as any model in Bedrock for seamless production deployment.

Automatic fault detection, diagnosis, and recovery without manual intervention. Continuously run model development workloads for months without disruption. Reduces training time by up to 40% through eliminated downtime.

Purpose-built Kubernetes extension for resilient foundation model training. Surgical recovery restarts only affected resources instead of full job restarts. Customizable hanging job monitoring via YAML configurations.

Train across hundreds or thousands of GPUs with built-in distributed training primitives. Data parallelism, model parallelism, and pipeline parallelism patterns supported out of the box with optimized communication.

Run JupyterLab, Code Editor, or connect local IDEs directly on HyperPod clusters. Interactive AI workloads on the same persistent EKS clusters used for training and inference.