Blitz cuts protein structure sampling to eight steps

When a structure predictor uses fewer refinement steps, does it preserve a molecule’s physical geometry as well as its similarity to a reference? A high unchecked similarity score can coexist with severe clashes or distorted bonds and angles, so the answer depends on checking both.

Blitz distills the diffusion modules of AlphaFold 3 and Boltz-2 into eight-step samplers, then uses geometry checks to assess the resulting structures.

The authors compare Blitz with its teacher on the primary Boltz-2 test set, where predicted structures are scored for local distance agreement and stereochemical defects.

OpenStructure measured checked complex lDDT of 0.8635 for the eight-step AlphaFold 3-Blitz student and 0.8615 for its 200-step teacher, with three recycle updates and five candidates for each.

Because this score includes checks for clashes and distorted bonds and angles, the close result supports the conclusion that this configuration retained teacher-level structural accuracy under the reported evaluation.

These results cover the tested models, datasets and scoring procedures. The reduction from 200 to eight steps applies to diffusion sampling; Blitz leaves the structure prediction trunk unchanged, so it does not establish the same speedup for the full prediction pipeline.