Six garnet pyroxenites from Beni Bousera, Morocco, yield a mean lutetium-hafnium age of 25 +/- 1 million years ago and show a wide range in hafnium isotope compositions (varepsilonHf = -9 to +42 25 million years ago), which exceeds that of known basalts (0 to +25). Therefore, primary melts of garnet pyroxenites cannot be the source of basalts. The upper mantle may be an aggregate of pyroxenites that were left by the melting of oceanic crust at subduction zones and peridotites that were contaminated by the percolation of melts from these pyroxenites. As a consequence, the concept of geochemical heterogeneities as passive tracers is inadequate. Measured lutetium-hafnium partitioning of natural minerals requires a reassessment of some experimental work relevant to mantle melting in the presence of garnet.