Quantifying the Capability Boundary of DeepSeek Models: An Application-Driven Performance Analysis