Exowatt beats frontier models on accuracy with SageMaker serverless model customization

This article describes how Exowatt used Amazon SageMaker AI serverless model customization to build task-specific fine-tuned models for alarm triage and energy-market analysis, achieving higher accuracy than frontier models at significantly lower cost.

• Fine-tuned Qwen3-1.7B model for edge alarm triage reached 0.911 macro-F1, outperforming Claude Sonnet 4.6 at 0.698
• Fine-tuned Qwen3-14B model for cloud energy-market analysis achieved 0.975 macro-F1 versus Claude Opus 4.7 at 0.625
• Edge model runs offline on CPU with no inference cost; cloud model costs $0.60 per 1,000 answers versus $29.26 for Claude Opus
• Serverless fine-tuning eliminated GPU provisioning bottlenecks, enabling 23 training jobs in six days
• Data quality and paraphrase diversity outweighed raw volume; 2,522 curated examples beat larger datasets
• Pipeline uses seven AWS serverless services with no instances, clusters, or endpoints to manage

Exowatt's reusable AI pipeline demonstrates that smaller, fine-tuned models can outperform large frontier models on specific tasks while reducing costs and enabling edge deployment without connectivity dependencies.