GoKawiilDeveloper fine-tunes open-source GLiNER model using Gemini-labeled data for $9
A hobbyist building a knife-tracking scraper for Reddit comments replaced expensive per-comment Gemini 3.1 Pro API calls with a fine-tuned version of the open-source GLiNER model. He had Gemini label 4,290 comments once, then used that labeled data to train GLiNER large v2.5 to extract knife brands, models and steel types, spending about $9 total on the labeling step.
GoKawiil's interpretation of the reporting above, not reported fact.
This is a concrete example of using an expensive frontier model as a one-time teacher for a small, free, locally-runnable model rather than paying for every inference call, a pattern known as model distillation that could apply broadly to narrow extraction tasks. It shows how a $9 one-time cost can eliminate ongoing per-request API bills while closing much of the accuracy gap that zero-shot open models have against paid large language models.
- Zero-shot GLiNER scored about 0.65 F1 versus Gemini on knife-related entity extraction, prompting the fine-tuning effort.
- Gemini 3.1 Pro was used only to generate training labels once, costing roughly $9 for 4,290 comments.
- The fine-tuned GLiNER large v2.5 model can now run for free, replacing ongoing paid API calls for this narrow named-entity recognition task.
Source: petervijeh.com, 2026-09-17
Published there as: “I had Gemini train its own replacement for $9”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.