Developer Shifts 50% of AI Workload to Local LLMs by Splitting Tasks by Role

Uehara, a solo developer at EarthLink Network Co., Ltd., reported on July 30, 2026 that local LLMs handled 50.3% of execution volume across his AI development pipeline, with cloud models accounting for 48.2%. Rather than selecting the most powerful model for all tasks, he divided work by role — routing lightweight classification jobs to smaller local models and heavier tasks to the cloud. Testing revealed that a 14B model matched a 72B model in classification accuracy and code generation pass rates, while the larger model ran five to six times slower. A memory miscalculation when co-hosting both models caused a four-and-a-half-minute stall, which was resolved by shortening the classifier's context window. Critically, high-risk operations involving databases, authentication, billing, and production systems are always routed back to human approval regardless of model output.