Smaller models (such as 0.8B) are suitable for edge devices or latency-sensitive real-time routing scenarios, while larger models (27B) are suitable for tasks requiring higher judgment accuracy