Fluent reading is characterized by rapid and accurate identification of words. It is commonly accepted that such identification relies on the availability of orthographic knowledge. However, whether this orthographic knowledge should be seen as an accumulation of word-specific knowledge in a lexicon acquired through decoding or as a well-developed associative network of sublexical units is still under debate. We studied this key issue in reading research by looking at the serial and/or parallel reading processes underlying word and nonword reading. Participants were 314 Dutch 2nd, 3rd, and 5th graders. The children were administered digit, word, and nonword naming tasks. We used latent class analyses to distinguish between readers who processed the letter strings serially or in parallel, based on the correlation patterns of word and nonword reading with serial and discrete digit naming. The 2 classes of readers were distinguished for both word and nonword reading. The validity of these classes was supported by differences in sensitivity to word and nonword length. Interestingly, the different classes seemed to reflect a developmental shift from reading all letter strings serially toward parallel processing of words, and later of nonwords. The results are not fully in line with current theories on the representation of orthographic knowledge. Implications in terms of models of the reading process are discussed.