Versatile models of human locomotion control can elicit new ideas for the control of legged robots and provide simulation test-beds for walking assistive robots. There exist neural control models that can generate human-like diverse and robust locomotion behaviors. However, most of these behaviors have been generated by extensive search on low-level control parameter sets, which is time consuming and limits the general applicability of the models. Our goal is to identify a hierarchical structure in neuromuscular control that allows to generate a large range of behaviors with a few high-level inputs. In this study, we focus on running. We incorporate a higher-layer speed adaptation policy to a previously proposed neuromuscular human model and find that it enables the model to run at speeds ranging from 2.4 to 4.0 ms−1 by changing only the target velocity. However, the speed changes occur slowly, and we investigate simple strategies that facilitate them. Among the strategies we explore, modulating the trunk lean shows fast and reliable acceleration and deceleration in average of 0.35 and −0.37 ms−2, respectively. The results show that the running speed of the neuromuscular model can be controlled to some extent with a higher-layer speed adaptation policy and a simple speed changing strategy. We plan to extend this framework to generate a larger range of locomotion behaviors with a few high-level commands.