This paper reports on an experimental study on the L3 acquisition of Mandarin temporal-aspectual sentence-final particles (SFPs) le, ne and láizhe by two groups of beginner-level learners (23 L1 English-L2 Mandarin learners and 30 L1 English- L2 Cantonese - L3 Mandarin learners). Both Mandarin and Cantonese are SFP languages while English is not. Aiming to shed light on the discussion on the source of transfer at L3 initial stages, we test predictions of three different L3 models by employing a sentence-picture matching task to elicit the learner’s interpretations of the target Mandarin particles. The results show that the L3 group was more accurate on the temporal meaning of le than the L2 group due to the help from the equivalent L2 Cantonese SFP, which rejects the prediction of the L1 factor hypothesis (e.g. Jin, 2009; Na Ranong & Leung, 2009) that transfer is always from the L1. Moreover, an acquisition asymmetry observed between le and láizhe suggests that additional factors such as syllabic mismatches, semantic transparency and word frequency can affect the initial transfer. In addition, the findings on the SFP ne indicate that typological proximity does not always facilitate L3A at the initial stage, if the previously learned languages are similar in the instantiation of the target property.