Exowatt beats frontier models on accuracy with SageMaker serverless model customization
aws ·
Exowatt has successfully utilized Amazon SageMaker AI serverless model customization to create specialized AI agents for its solar energy infrastructure, outperforming frontier models in accuracy while significantly reducing costs. The company developed two distinct fine-tuned models: a Qwen3-1.7B variant for edge-based alarm triage and a Qwen3-14B variant for cloud-based energy market analysis. By leveraging serverless training, Exowatt eliminated the need for manual GPU provisioning, allowing them to iterate rapidly and deploy production-ready models within days rather than months. The edge model operates offline on existing CPU hardware, achieving a macro-F1 score of 0.911 without connectivity or inference costs. Meanwhile, the cloud model delivers a macro-F1 of 0.975 at approximately 49 times lower cost per thousand answers compared to leading proprietary models like Claude Opus 4.7. This approach demonstrates that smaller, task-specific fine-tuned models can surpass larger general-purpose models when trained on high-quality, curated data. The resulting pipeline provides a repeatable path from raw operational data to certified deployment artifacts across both constrained edge environments and elastic cloud workloads.
aws