The acquisition of mathematical literacy in primary school is a complex process that is influenced by a large set of variables. A multilevel model was applied to identify significant predictors of mathematics performance in Chinese primary schools. Data were obtained from 10,959 students of six grades from primary schools in rural/urban areas, within five provinces with different developmental levels. At the school level, the aggregated socioeconomic status of a school is a significant predictor of math performance (x 2=4.3, df=1,p<.05), until the individual reading level is included. At the class level, grade is a significant predictor. Teacher's graduation level predicts performance (x 2=4.84, df=1,p=.03), until individual students' metacognition level is added. At the student level, reading performance (x 2=434.87, df=1,p<.00), mathematics self-efficacy (x 2=392.62, df=1,p<.00) and metacognition (x 2=756.62, df=1,p<.00) play a large and significant role. Socioeconomic status of family is a weak and polynomial predictor. The results reveal that individual background variables are important predictors and explain 46.67% of the total variance in math performance. After controlling for student characteristics, school and class level variables disappear as predictors, implying an interaction between contextual and individual variables. The present research findings have - next to theoretical implication - also policy implications for Chinese mathematics education. Firstly, the educational quality between provinces seem to be balanced, but the school quality within a province does not seem to be balanced. Secondly, there seems to be a need for a quality control related to the output of open teacher training institutions. Thirdly, remedial or intervention programs have to be put in place, to be proactive as to difficulties of students with different language backgrounds.