In scientific inquiry learning, manipulatives have been widely utilized as learning resources. Studies have explored the advantages of virtual manipulative (VM) for conceptual understanding and knowledge construction in science education. However, research on the mental engagement and perception of students during collaborative learning under different modalities of using VM remains rare. In this study, we designed a virtual lever manipulative (VLM) and three modalities of using VLM in a primary science course: one VLM per student, one VLM per group, and one VLM per class. There were 80 fifth graders from three classes who participated in this quasi-experimental study. They were asked to complete a group worksheet during collaborative learning activities. Cognitive load, as well as flow experience, was invested through a questionnaire survey after the learning activities. Task involvement was calculated by using the mental effort dimension of cognitive load and post-test scores. The findings indicate that class B (one VLM per group) gained the highest scores in group worksheet and the post-test 1 week later, followed by class C (one VLM per class) and class A (one VLM per student). Class B had the highest level of task involvement, as they had a shared screen among group members. In contrast, class A had a relatively low task involvement and spent more time checking consistency. Besides, both classes A and B had a higher sense of flow experience than class C. Class C experienced a traditional teaching method and less interaction with learning materials, leading to a lower level of flow experience and moderate task involvement.