Small AI models outperform larger ones in extracting job posting data

Job Opportunities API (JOA) tested whether small AI models could match large models at extracting structured data from job postings. They created an evaluation arena using 250 real job listings and scored 68 open-weight models on extracting 11 specific fields. The top-performing model, Qwen3.5-27B, scored 89.2 in initial testing but required thousands of tokens per listing. After fine-tuning three smaller models using Qwen3.5-27B's outputs, the smallest model (granite-4.2-3b) achieved the highest score of 93.3 while using only 117 tokens per answer. The results demonstrate that task-specific fine-tuning enabled small models to outperform much larger ones for this specialized extraction task.