This study investigated how the use of cohesive devices predicts writing fluency for second language (L2) undergraduate students (N = 99). Linear mixed effects models were built to predict writing fluency using cohesion indices. Results showed that the use of semantic overlap between adjacent sentences negatively predicted fluency in process. Furthermore, the use of more unattended demonstratives related to higher fluency in process but greater revisions, whereas more attended demonstratives associated with fewer revisions.