1.7B Model Outperforms 744B GLM-5 on Schema-Guided Dialogue — 437x Size Gap

1.7B Model Outperforms 744B GLM-5 on Schema-Guided Dialogue — 437x Size Gap

New findings from @j_golebiowski show a 1.7B parameter model surpassing GLM-5 at 744 billion parameters on the Schema Guided Dialogue benchmark — a 437-fold parameter-size disadvantage reversed. Notably, the smaller model maintains its advantage even when trained on corrupted data, suggesting the task-specific training signal is robust against noise.

Why It Matters

This finding reinforces the emerging pattern that task-focused small models can decisively outperform general-purpose frontier models on structured tasks. For practitioners building dialogue systems, it points toward smaller, cheaper, faster models as the practical deployment choice over frontier-scale generalists.