Small LLMs Democratize AI for Local, Specialized Workloads.

AI democratization hinges on the ability of institutions to select, audit, and specialize capable models under realistic hardware and governance constraints, rather than solely on matching frontier-scale generality. This research investigates this challenge by evaluating nine open-weight language models, ranging from 135 million to 3 billion parameters, on a custom 1,085-example, 16-topic multiple-choice benchmark. The benchmark specifically targets symbolic precision, constrained formatting, extraction, and short-horizon semantic decision-making, with a strict one-letter output protocol. A shared parameter-efficient fine-tuning (PEFT) pipeline, utilizing 4-bit NF4 quantization with DoRA/LoRA-style adapters, was applied to a subset of these models within an NVIDIA L4-class budget. In base evaluations, Qwen Coder 3B performed best at 75.67% accuracy, followed by Qwen2.5 1.5B and Granite 3.3 2B. Post-adaptation on a held-out fine-tuning split, models like Qwen Coder 3B saw significant improvements, gaining over 26 points in accuracy. The findings consistently show that a disciplined workflow involving benchmark construction, cross-model evaluation, and low-cost specialization makes sub-3B models viable as local experts for structured niche workloads. This suggests that powerful AI capabilities can be deployed without relying on massive, expensive frontier models.