We present the first Quantum-Natural-Language-Processing (QNLP) proof of concept for token-level language identification in Martinican Creole, a low-resource language with substantial lexical overlap with French. Our approach compares: (1) a character-bigram baseline, (2) a classical logistic regression model trained on three interpretable intra-token features (entropy, vowel ratio, syllable complexity), and (3) a compact three-qubit variational circuit embedding the same features in product-state and entangled configurations. On token-level macro- \(F_{1}\) and accuracy, the classical model performs slightly better than the entangled variant, reflecting its robustness on high-frequency items where individual features are most discriminative. However, a type-level analysis shows a different pattern: the entangled embedding reduces the number of distinct misclassified types and uniquely corrects a sizable subset that both the classical and product-state models miss. The rescued set is heterogeneous (i.e., mixing frequent Creole items and proper names with cross-lingual orthography), and some gains are dataset-contingent serendipitous effects, since the models are context-blind and rely only on intra-token features. Our findings demonstrate that even minimal quantum embeddings can capture non-linear interactions between interpretable features in ways that simple linear models and product mappings cannot. The results highlight the potential of hardware-efficient QNLP for fine-grained, low-resource language identification, especially in settings where context is unavailable but token-internal cues are strong. We conclude with a roadmap for extending this work to richer feature sets, context-aware hybrid architectures, and cross-lingual transfer to other Creole and closely related language pairs.