Exowatt's fine-tuned Qwen3 models beat frontier models at up to 49× lower cost
AWS ML Blog · rss · 2026-10-28
AWS ML Blog case study: solar company Exowatt used SageMaker AI serverless model customization to build a repeatable fine-tuning pipeline whose task-specific small models beat frontier models at up to 49× lower cost. Their P3 solar systems power AI data centers, and thousands of field sites need edge alarm triage — frontier models failed on narrowness, no-GPU/no-internet edge constraints, and per-call economics. The stack: serverless LoRA+SFT fine-tuning from SageMaker Public Hub recipes enabling multiple experiments per day; Qwen3-1.7B quantized to GGUF and run via llama.cpp on no-GPU edge controllers offline; Qwen3-14B served via Bedrock Custom Model Import for cloud energy-market analysis with no idle cost; paraphrasing via Qwen3-VL-235B and Claude Sonnet 4.6; serverless evaluation scoring 2,000 answers per candidate. The durable output is the pipeline itself — raw ops data to certified training set to evaluated deployment in days instead of months.
More from coding & agent
- Dev builds AI-native macOS dock where an agent classifies activity and customizes focus modes in real time — addyosmani · 2026-10-28
- ACARE: A 4B Reranker Scores Documents Against the Agent's Full Reasoning Context — _reachsumit · 2026-10-28
- Salesforce launches CRM-Agent-Bench: contract-based benchmark for agents in enterprise environments — salesforceaicloud · 2026-10-28
- Judgment Labs opens general access: monitoring and eval layer for long-horizon agents — Scobleizer · 2026-10-28
- Clay's Claygent handled 5B+ queries, cut inference costs 47% with a model router — rhythmrg · 2026-10-28
- Kaggle co-founder finds Claude Code's .claude/worktrees quietly ate 449GB of his disk — benhamner · 2026-10-28