Multiword combinations perform a crucial role in signifying fluency, accuracy and idiomaticity in academic writing. Lexical bundles are recurrent, but not salient, multi-word combinations, for example, on the other hand, the fact that the, and it should be noted. They are important as they act as discourse frames to relate to new information or as interactional devices to mark the involvement of the writer and the reader. These functions can also be regarded as metadiscoursal functions, represented by metadiscoursal models. The use of lexical bundles in L2 academic writing has been the focus of a number of recent studies, but few studies distinguish bundles in different sentence positions, investigate bundles from the perspective of metadiscoursal functions, and explore the reasons underlying the bundle choices of L2 writers. The present study sought to fill these gaps by comparing the use of sentence initial bundles (i.e. bundles at the beginning of sentences) in Chinese L2 and New Zealand L1 thesis writing in the discipline of general and applied linguistics. Four collections were built: a Chinese masters thesis corpus, a New Zealand masters thesis corpus, a Chinese PhD thesis corpus and a New Zealand PhD thesis corpus. In comparing these four corpora, this study provided a detailed picture of the use of sentence initial bundles in Chinese postgraduate writing and an overall picture of variation in bundle use across different postgraduate levels of students in terms of frequency, structure and function. Semi-structured interviews with six Chinese postgraduates were conducted after the text analysis to understand the reasons for Chinese students’ bundle choices. The interviews were based on the expressions in participants’ original drafts, which were completely or partially overlapped with the sentence initial bundles generated from the corpus data. Chinese masters and PhD students were found to rely more heavily on sentence initial bundles, particularly interactive bundles. They preferred to start sentences with PP-based bundles, VP-based bundles, and conjunction + clause fragment bundles; but were less aware of the importance of NP-based bundles and anticipatory-it bundles. With regard to function, both the Chinese PhD and masters corpora were characterised by a heavy use of condition bundles and booster bundles; and a relatively low use of endophoric bundles, attitude bundles, hedge bundles, self-mention bundles and directive bundles of cognitive acts. In regard to bundle development, both groups of masters students were found to use more bundles than their PhD counterparts. However, the two PhD groups shared more bundles. More research-related NP-based bundles occurred in masters corpora, and more PP-based bundles and anticipatory-it bundles appeared in PhD students’ writing. A functional analysis showed that both groups of PhD students used more transition bundles, condition bundles, section-level frame bundles and self-mention bundles, but fewer attitude bundles. Interviews with six Chinese postgraduates revealed possible reasons for Chinese students’ bundle selection and use, which included but were not limited to interlingual transfer, classroom learning, noticing in reading, a lack of rhetorical confidence, and misunderstanding of rhetorical conventions. The findings suggest the need to go beyond the teaching of lexical bundles as a list of fixed multiword expressions. Teachers and learners are advised to address the pedagogical implications of bundle studies, and to use corpus-based tools (e.g. FLAX) to approach bundles as lexico-grammatical frames in which slots can be filled with a variety of words.