Exowatt says SageMaker-customized Qwen3 models surpassed frontier-model baselines on energy alarm triage and market analysis at substantially lower inference cost.
Exowatt says it used Amazon SageMaker AI’s serverless model-customization tools to build two task-specific models for its energy operations: Qwen3-1.7B for edge-based alarm triage and Qwen3-14B for cloud energy-market analysis. In an AWS blog post dated October 27, 2026, the companies reported that both models scored higher on their respective evaluations than the larger general-purpose models used as baselines.
For alarm triage, the 1.7-billion-parameter model reached a macro-F1 score of 0.911 on 320 held-out cases, compared with 0.698 for Claude Sonnet 4.6. After 8-bit quantization, AWS says it ran on an existing site CPU in a 6 GiB llama.cpp environment, without a GPU or internet connection. The model’s training set contained 2,522 examples, combining 684 verified decisions with certified paraphrases. Training took about 32 minutes and cost approximately $20–$25, according to the post.
The cloud model classifies energy-market events from live market data. Exowatt reported a 0.975 macro-F1 score for Qwen3-14B on 2,000 test cases, versus 0.625 for Claude Opus 4.7. The post says the customized model cost $0.60 per 1,000 answers, compared with $29.26 for the baseline—about 49 times less. The model was trained using historical Electric Reliability Council of Texas data; the team expanded labeled examples with paraphrases while keeping labels and numerical fields fixed.
AWS and Exowatt attribute the results to narrowly defined tasks, curated data and fine-tuning rather than model size alone. These are company-reported evaluations on Exowatt-specific workloads, not a general comparison across tasks. The work also produced a repeatable pipeline for preparing data, training models and evaluating deployment versions. Exowatt says the process can run in days rather than months and plans to apply it to additional operational tasks as its energy-system fleet grows.