A knife-brand extractor cost $11.50 to build

Peter Vijeh spent $11.50 total to build a Reddit knife-brand extractor, replacing per-call API pricing, according to his GLiNER fine-tuning write-up. The split was $9 for Gemini labeling and $2.50 for GPU time. After that, the model runs locally with no per-call cost, no matter the volume.

He had Gemini 3.1 Pro label 4,290 Reddit comments once, tagging three attributes: knife brands, models, and steel types. Character offsets came from code, not manual annotation, so the labeling pass covered the full dataset without per-example review.

That labeled set trained GLiNER large v2.5, a 459M-parameter model, on a single Tesla T4 GPU. The fine-tune scored 0.83 F1 on held-out validation, with brand extraction at 0.904 F1 and material recall at 0.911.

The catch is the schema has to hold still. Three fixed attributes made one labeling pass enough to cover the space. A schema that keeps changing means relabeling and retraining every time, and the economics stop working.