Several models have been proposed for the structure of amyloid fibrils formed from intact β2m and its peptide fragments. A solid-state NMR study of a 22-residue peptide revealed pairs of β-strands packed in a parallel, staggered arrangement.39 A zipper-spine model has been proposed for fibrils formed from the intact protein, with residues 83–89 forming a central β-spine and with the rest of the sequence postulated to remain in a native-like state.40 Microcrystals of a short peptide (82NHVTLS87) show a simple cross-β ribbon, similar to the structures found in other peptide crystals of these presumed amyloid mimetics.41 Models based on crystal structures of β2m dimers and hexamers and others incorporating domain swapping or stacking of native subunits have also been proposed.42–45 None of these models is compatible with the 3D structure of the β2m amyloid presented here, a key feature of which is the globular dimer-of-dimers repeat. Consistent with the expectation of a non-native protein fold in the fibril, the density representing a dimer of β2m subunits (roughly 4 nm × 3 nm × 3 nm) cannot be accounted for by two copies of the native monomer (∼ 4 nm × 2 nm × 2 nm),46 stacked either axially or side by side, nor does a previously observed crystallographic dimer fit this density.45 This lack of fit is demonstrated by docking the atomic structures of the crystallographic dimer of the variant P32A,45 a domain-swapped dimer derived from molecular dynamics simulations,44 and a simple cross-β model or four native β2m monomers into the EM density of an extracted dimer-of-dimers unit (Supplementary Fig. 3). One way that the native-like monomers could fit is to be in staggered pairs (Supplementary Fig. 3e). However, in such an arrangement, interfaces completely different from those previously proposed or observed are required in order to fill the density, in addition to significant rearrangements of at least the surface loops to optimise packing and to avoid steric clashes. Moreover, such an arrangement does not yield the well-organised cross-β assembly expected in amyloid fibrils9 and observed for the fibrils analysed here29 and is incompatible with the proposed parallel arrangement of β-strands in these fibrils as indicated by FTIR.31 While the structure of β2m monomers within the fibril remains unresolved, refolding from the initially highly disordered monomeric state at pH 2.526 upon amyloid assembly results in a globular, but non-native structure. This conformation is restrained by the native disulfide bond, which is required for fibrillation of monomeric β2m in vitro and in vivo.22