Lithium isotopes are increasingly used as a tracer for the sources of mafic-ultramafic intrusive rocks, but the validity of such approach has never been evaluated directly. We have addressed this issue by analyzing mineral separates from an arc-type mafic-ultramafic complex (namely Xiarihamu) in the Tibetan plateau for Li isotopes. The Xiarihamu complex consists of a gently-dipping, ultramafic chonolith in the center and older gabbro in the margins. The chonolith was interpreted to be a feeder for olivine-charged magma by previous studies. Our new data show minor to strong isotopic disequilibrium between coexisting minerals in the chonolith, with lower delta Li-7 in clinopyroxene than in coexisting orthopyroxene by up to 7 parts per thousand. The total range of delta Li-7 in the pyroxene and olivine separates together is from 5.0 parts per thousand to -16.2 parts per thousand, which is significantly different from that of known arc lavas in the world (3.6 +/- 1.2 parts per thousand). Samples from a subvertical drill core of similar to 190 m in length penetrating the chonolith show a concentric decrease of delta Li-7 in orthopyroxene from -16.2 parts per thousand in the middle to +1.5 parts per thousand and + 3.8 parts per thousand in both ends. The delta Li-7 values of mineral separates show no correlation with whole-rock Sr-Nd isotopes and incompatible trace element ratios, as well as olivine forsterite contents and orthopyroxene Mg/(Mg + Fe) molar ratios, indicating that the observed great variability of delta Li-7 in the minerals was not caused by variable subduction inputs or magma differentiation alone. A positive correlation exists between orthopyroxene delta Li-7 and whole-rock LOI, illustrating that serpentinization-talc alteration caused delta Li-7 in the residual mineral to increase, contributing to the great variability of delta Li-7 on the mineral scales. Subducting sediment input during magma generation and crustal contamination together cannot adequately account for the observed extremely low delta Li-7 (-2 to -16.2 parts per thousand) in the minerals from many of the rock samples as well as the unusually low average delta Li-7 (-3.5 parts per thousand) of all of the samples, because the delta Li-7 of the crust and subducting sediments are much higher. Kinetic fractionation of Li isotopes in response to a temperature gradient across the chonolith as well as to subsequent diffusion of Li from interstitial melt to the mineral phases is the best explanation. Since the kinetic fractionation is independent of tectonic settings, such effect is inevitably present in all mafic-ultramafic intrusions on Earth, illustrating that the effect of kinetic Li isotope fractionation must be considered before using Li isotopic data to constrain mantle sources and magmatic evolution, especially if mineral rather than whole-rock data are employed.