Solar-energy company Exowatt has built two task-specific Qwen models for operational work that previously depended on larger general-purpose systems. One model handles equipment alarms on low-power edge controllers, while the other analyzes energy markets in the cloud.
The edge system uses Qwen3-1.7B because sites may have intermittent connectivity and no GPU. A separate Qwen3-14B model serves the more demanding cloud task. Exowatt used SageMaker’s serverless fine-tuning and evaluation services to turn operational data into training sets, test 2,000 responses per candidate and register approved artifacts for deployment.
AWS says the resulting models achieved higher accuracy than the frontier models used for comparison at up to 49 times lower cost. The claim is tied to two narrow, proprietary tasks and an AWS customer case study; it does not mean small models generally outperform frontier systems. Evaluation quality also depends on how Exowatt defined correct alarm actions and market-analysis answers.
The reusable part is the pipeline rather than either model. Exowatt says it can now move from raw data to a certified training set and evaluated deployment artifact in days instead of months. Narrow models can be attractive when a task repeats often, data is specialized and offline operation matters. They still require monitoring as equipment, alarm patterns and business rules change.