Generative modeling for single-cell RNA-seq has proven transformative in crucial fields such as learning single-cell representations and perturbation responses. However, despite their appeal in relevant applications involving data augmentation and unseen cell state prediction, use cases like generating artificial biological samples are still in their pioneering phase. While common approaches producing single-cell samples from noise operate in continuous space by assuming normalized gene expression, we argue for the necessity of sample generation in a raw transcription count space to favor processing-agnostic data generation and flexible downstream applications. To this end, we propose cellFlow, a Flow-Matching-based model that generates single-cell count data. In our empirical study, cellFlow performs on par with existing methods operating on normalized data when evaluated on three biological datasets. By carefully considering raw single-cell distributional properties, cellFlow is a promising avenue for future developments in single-cell generative models.