Hypergraph neural networks (HNNs) have recently emerged as a promising paradigm for modeling higher-order relations, yet their study under heterophilic settings remains highly limited. Empirical evidence shows that classical message-passing–based HNNs suffer severe performance degradation on hypergraphs with low homophily ratios. We attribute this primarily to heterophily mixing, where semantic signals from neighbors of different classes become entangled, thereby eroding the discriminative power of node representations. To address this challenge, we propose HyperUnmix, a novel method that disentangles heterophily mixing through a mixed-curvature manifold. Guided by the intuition that nodes of different classes exhibit distinct distributional characteristics, we model the representation space as a Cartesian product of multiple hyperbolic submanifolds, each aligned with a specific class. By constraining information flow to propagate mainly within the submanifold corresponding to its class, HyperUnmix effectively alleviates mixing during aggregation. Extensive experiments on both heterophilic and homophilic hypergraph benchmarks demonstrate that our model establishes new state-of-the-art performance, providing fresh insights into heterophilic hypergraph learning. The source code is available at https://github.com/liulizhi1996/HyperUnmix .