AlphaFold 3 v3.0.4 dropped — and I’m credited in the release notes for a memory optimization merged into this version.

The headline feature is that AlphaFold 3 is now runnable on CPU-only machines and Apple Silicon, which meaningfully expands access beyond GPU clusters.

My contribution (PR #674) is in there: the OuterProductMean module was rewritten to replace two sequential einsum operations with a single fused three-way einsum, cutting peak intermediate tensor memory from ~256 MB to ~72 MB — a 3.56x reduction with no throughput regression.

Other highlights in v3.0.4: Google Cloud Storage path support, chain IDs added to summary confidence JSON, JAX updated from 0.9.1 to 0.10.2 for faster inference and Blackwell GPU support, and an early validation fix for RNG seed type checking.

Good to see it ship.