Why it exists

Large frontier models (like GPT-4) are slow and expensive. Distilled student models deliver 80-90% performance at a fraction of latency and cost.

How it works

The student model is trained on completions and logit distributions generated by the teacher model across thousands of representative prompts.