Exowatt beats frontier models on accuracy with SageMaker serverless model customization

Exowatt used Amazon SageMaker AI serverless model customization to fine-tune small open-weight models for its solar energy systems, saying they beat frontier models on accuracy at up to 49× lower cost.

• The company needed an agent that could triage alarms on its energy management system without a GPU or an internet connection, since many sites have limited connectivity.
• Exowatt fine-tuned Qwen3-1.7B for edge alarm triage and Qwen3-14B for cloud energy-market analysis, and plans to reuse the pipeline for smaller task-specific models across its fleet.
• The edge model is frozen, quantized to GGUF and run with llama.cpp inside Kubernetes pods on on-site controllers, requiring no GPU and no internet.
• The cloud model is served through Amazon Bedrock Custom Model Import, which runs with no idle cost and bills only while the model is active.
• Exowatt says the durable result was the pipeline itself, taking days rather than months, with the two models as its first outputs.