Diffusion-based generative models for speech enhancement often face challenges in balancing performance and inference efficiency. We propose a Velocity-guided Interpolant Diffusion Model for Speech Enhancement (VISE), which incorporates three key innovations: a scalable interpolant framework that reconstructs the reverse diffusion process using velocity terms and state variables, a loss function designed to fit the velocity terms for efficient data distribution learning, and a combined SDE/ODE sampling strategy with an optional corrector. Experiments on VoiceBank-DEMAND and WSJ0-CHiME3 datasets show that our VISE significantly outperforms baselines across multiple metrics, particularly in noise separation with SI-SIR improvements of up to 4.7 dB. Moreover, its inference efficiency is up to 7 \(\times \) faster than existing diffusion-based methods while maintaining excellent enhancement performance.