It is well documented that grouping smaller units of information into larger units (i.e. chunking) can help offset working memory (WM) capacity limit. A long standing question is whether working memory capacity limit is determined by the sheer number of chunk-based units, or memory resources are flexibly distributed over multiple chunks depending on tasks and also the resolution level at which memory representations are encoded. Results from previous work in visual WM has shown evidence for both positions. The current work explores the effect of rational resource distribution for storing linguistic information in verbal WM. Two experiments were conducted using a change detection task and Mandarin Chinese words as stimuli. The task in Experiment 1 encouraged the encoding of the lexical representations at relatively lower granularity level, and the task in Experiment 2 targeted more detailed representation of a word (higher granularity). Two findings were noteworthy. First, compared to the memory performance of non-words, the memory performance of words showed clear benefits of chunking. Second, memory performance under more precise encoding (Experiment 2) was worse than under less precise encoding (Experiment 1), but only when memory load was also high. Our findings lend support to the rational resource models that allow dynamic distribution of limited memory resources based on the demand for representational granularity.