Large datasets generated by Illumina sequencing are ideally suited to transcriptome characterization. We generated 3,052,501 27-mer reads from F1 mouse embryonic stem (ES) cell cDNA. Using the ELAND alignment tool, 74.5% of reads matched sequenced mouse resources, <1% were contaminants, and 3.7% failed quality control. Of the reads, 21.6% did not match mouse sequences using ELAND, but most of them were successfully aligned with mouse mRNAs using MegaBLAST. We conclude that most of the reads in the dataset are derived from mouse transcripts. A total of 14,434 mouse RefSeq genes were represented by at least 1 read. A Pearson correlation coefficient of 0.7 between Illumina sequencing and Illumina array expression data suggested similar results for both technologies. A weak 3' bias of reads was found. Reads from genes with low expression had lower GC content than the corresponding RefSeq genes, indicating a GC bias. Biases were confirmed with further Illumina read datasets generated with cDNA from mouse brain and from mutagen-treated F1 ES cells. We calculated relative expression values, because transcript length and read number were correlated. In the absence of signal saturation or background noise, we believe that short-read sequencing technologies will have a major impact on gene expression studies in the near future.