Beating Frontier Models on a Turkish Classification task for $30 of GPU + RL
Towards AI
•
Machine Learning
Generative AI
AI Hardware
AI Research
Reinforcement Learning
Last weekend I got inspired and post-trained a small Turkish model for e-commerce attribute extraction. It beat Opus 4.7, GPT-5.5, and Gemini 3.1 Pro at 1,635× lower inference cost. Here’s what I learned, and how you can do something similar cheaply if you have nothing to do this weekend. Pipeline: scrape Trendyol (Amazon of Turkey) products, score with verifiable rewards, train Asure-12B with GRPO, evaluate against frontier APIs Most teams reach for frontier APIs by default. That is usually the right move. Frontier models are broad, robust, and convenient.