The same numerical magnitude can be manifested in different physical notations. However, how the numbers with distinct notations are mentally represented is still unclear. Here, we hypothesized that how the number is mentally represented is influenced by the numerical processing level of the tasks. If the task only needed a low-level processing, the representation would be dependent on the surface forms of the numbers, exhibiting a numerical notation-dependent effect. By contrast, if the task required a deeper magnitude processing, the processing would utilize an abstract numerical format whose effects are notation independent. To test this hypothesis, we manipulated the notation type and the numerical processing level of the tasks. An ERP component N270 was taken to index the mismatch between the internal representation and the probed number. The results showed that N270 was enhanced when the magnitude was mismatched between two numbers. More importantly, under the task requiring a low-level processing (e.g., magnitude comparison), compared with the same notations, the latency of N270 difference wave was delayed by different notations, exhibiting a notation-dependent effect. However, in the task involving a deeper processing (e.g., magnitude addition), the N270 latencies were earlier for probes having distinct notations (Mandarin or Arabic number) than for probes having the same notations as in the addition operation (Mahjong). Moreover, no difference was found on N270 latencies between the two distinct notations but with similar degree of familiarity. Taken together, these results support our hypothesis that the numerical processing level affects the number representation.