A language learning platform just swapped expensive frontier AI prompting for a small, fine-tuned model, and grammar tracking got better and 16 times cheaper.
Researchers fine-tuned Qwen3.5 small language models on filtered, rebalanced training data generated by larger teacher models, then deployed a compact 0.8B-parameter version to track grammar mastery for every English learner on the platform. The system reads learner-tutor lesson transcripts and tags grammar concepts, evidence spans, and correctness judgments, turning scattered in-lesson corrections into a running mastery score. On two human-curated benchmarks, both the deployed 0.8B model and a larger 4B reference version beat prompted GPT-5.4 and GPT-5.6 Sol on precision and recall, even as the matching criteria got stricter. Serving the small model costs roughly 16 times less than prompting a frontier model for the same job.
This is a case study in a trend bigger than one language app: a small model trained on a narrow, well-defined task can beat a general-purpose frontier model doing that same task through prompting, at a fraction of the inference cost. It suggests the economics of AI tutor features depend less on model size than on whether a company bothers to collect and curate task-specific training data. The business numbers back it up: the platform reported 15.8% higher learner engagement, a 2.1% bump in scheduled lesson hours, and 13.2% more gross bookings (GMV) from new lessons after rollout.
Fine-tuning a cheap model for one job beats renting a famous one to do it adequately.