Exowatt Fine-Tunes Small Models via SageMaker to Outperform Frontier AI
Frontier AI models · Enterprise AI · Smart operations
Exowatt adopted Amazon SageMaker AI serverless customization to fine-tune compact open-weight models for edge alarm triage and energy-market analysis, achieving higher task accuracy than frontier models at lower operational costs.
Exowatt, an energy systems company powering data centers, implemented Amazon SageMaker AI serverless model customization to create task-specialized AI models. By fine-tuning Qwen3-1.7B for offline edge controllers and Qwen3-14B for cloud market analysis, the firm avoided dedicated GPU management while outperforming frontier general-purpose models on domain tasks.
Key takeaways
- Qwen3-1.7B fine-tuned via SageMaker reached 0.911 macro-F1 in edge alarm triage, running on on-site CPUs without GPUs or internet.
- A fine-tuned Qwen3-14B attained 0.975 macro-F1 on ERCOT energy-market analysis, delivered via Bedrock Custom Model Import at $0.60 per 1,000 answers.
Source：AWS Machine Learning Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.