The learnware paradigm aims to help users leverage numerous existing high-performing models instead of starting from scratch, where the learnware consists of a well-trained model and the specification describing its capability. Generally, previous studies assumed shared feature spaces for all models and user tasks, but in reality, feature spaces often differ. If the learnware paradigm can handle this scenario, \textit{many heterogeneous models built for a task can be identified and reused for a new user task}. In this paper, we find that \textit{label information}, including model prediction and user's minor labeled data, is crucial and previously unexplored. We explicitly explore this information to address the problem. This paper proposes a new specification that encodes model capabilities and enhances subspace learning for better learnware management. It also recommends considering the conditional distributions with label information to improve learnware recommendations. Extensive experiments show that the recommended heterogeneous learnware significantly outperforms user self-training with limited labeled data and continues to enhance performance as more labeled data becomes available.