This article reports on an empirical study of L3 Mandarin, aiming to shed light on transfer effects and their interaction with other factors throughout the L3 acquisition trajectory. A fill-in-the-blank task was employed to examine L2 and L3 acquisition of three types of Mandarin sentence-final particle clusters. Participants in the study were Mandarin native speakers, L1 English-L2 Mandarin speakers, L1 English-L2 Cantonese-L3 Mandarin speakers, and L1 Cantonese-L2 English-L3 Mandarin speakers at low and high Mandarin proficiency levels. Our data support arguments in the L3 literature that transfer at the L3 initial stages is from a structurally more similar language. Facilitation takes place when the instantiation of the target property in the L3 is exactly the same as that in the transfer source language. Overall typological similarity might be helpful when transfer is from the L1, but it does not always facilitate the L3 acquisition of individual properties. For later L3 development, construction frequency and mapping paradigms have a great impact on the acquisition outcome. Moreover, a certain kind of non-facilitative transfer from an L1 is found to persist in L3 Mandarin grammars, which qualitatively supports the Cumulative Input Threshold Hypothesis.