Exowatt introduces a repeatable model customization workflow for operational alarms and the energy market

Exowatt used Amazon SageMaker AI serverless model customization to build a repeatable workflow covering operational data preparation, fine-tuning, evaluation and deployment of specialized models. It uses Low-Rank Adaptation (LoRA) and supervised fine-tuning (SFT). It stores data and the resulting models in Amazon S3, evaluates 2 000 responses from each candidate model and records the results in SageMaker Model Registry.

For operational alarm classification, the company customized Qwen3-1.7B. After quantizing it to GGUF format, it runs the model using llama.cpp on local control devices without GPUs or an internet connection. For energy market analysis, it used the same workflow with Qwen3-14B, which it deploys through Amazon Bedrock Custom Model Import. According to the service description, only active model usage is billed, with no charges for idle time.

According to AWS, the workflow allows model customization to be completed in days instead of months and delivers higher accuracy than leading models on the tasks described, at up to 49× lower costs. These results are reported in a case study by the service provider. Details are available in the source article.