Recent text‑to‑motion models have achieved impressive motion quality and instruction following, yet many‑step denoising and large model components make deployment slow and memory‑intensive.\ \ We introduce Terminal‑Amplification‑Controlled Distillation (TACD), an on‑policy approach that trains efficient motion generators from text prompts and pretrained teachers without any real‑motion data. TACD supervises clean‑motion predictions along student‑generated trajectories using segmented on‑policy flow distillation.\ \ A failure mode appears when velocity matching on a fixed supervision grid repeatedly over‑weights errors near the denoising endpoint, degrading few‑step generation. TACD ties the latest teacher query to the student’s step size, bounding effective loss weights in clean‑motion space while keeping inference unchanged.\ \ Experiments on HumanML3D and KIT‑ML show that an eight‑step student reduces FID by 58% compared with distillation without this bound. For diffusion teachers, the endpoint‑matching variant of TACD yields four‑step students with lower FID and text‑motion retrieval comparable to or better than their 50‑step teachers.\ \ On HY‑Motion and Kimodo, eight‑step students achieve 7.7‑11.9× end‑to‑end speedups and cut peak GPU memory by 3.8‑6.7×.\ \ Project page: https://vkgo.github.io/TACD/\ \ Review: TACD’s dynamic loss‑weight bounding substantially improves few‑step generation quality and inference efficiency, offering a practical path toward lightweight deployment of text‑to‑motion systems.