ByteDance's SeedFold outpaces AlphaFold 3

SeedFold is a biomolecular structure prediction model from ByteDance's Seed team, designed to push the boundaries of protein and complex structure modeling through scaling.

Alongside the full SeedFold, the team introduced SeedFold-Linear, a high-efficiency variant with linear triangular attention. Across the FoldBench benchmark, both models achieve state-of-the-art results, beating AlphaFold 3 on the majority of protein-centric tasks.

How they pulled it off

Instead of making subtle architectural tweaks, the researchers attacked the model capacity bottleneck from three distinct angles.

First, they scaled up width. The Pairformer dimension was widened from 128 to 512—a 4x increase in width. The team identified pair representation as the key structural bottleneck, finding that wider representations directly boost the model's ability to capture complex biological interactions, even if it introduces serious memory and training stability hurdles.

Second, they replaced vanilla triangular attention. Standard triangular attention runs at O(n³) computational complexity. The team engineered two linear triangular attention alternatives that drop the complexity down to O(n²). One blends vanilla strengths with linear attention designs, while another adds a gating mechanism that excels on DNA and RNA prediction tasks.

Third, they used heavy distillation. Since experimental structural datasets are notoriously scarce, the team distilled 26.5 million samples using AlphaFold 2. That effectively expanded their training corpus by 147x compared to available experimental structures alone.

Benchmark performance

The FoldBench numbers back up the engineering choices.

On antibody-antigen (Ab-Ag) modeling, SeedFold hits a 53.21% success rate, comfortably ahead of AlphaFold 3 at 47.90%, Boltz-1 at 33.54%, and Chai-1 at 23.64%.

For monomer prediction (lDDT 0.8889 vs. 0.88) and protein-protein docking (74.03% vs. 72.93%), SeedFold holds a steady edge over AlphaFold 3. In protein-ligand tasks, SeedFold-Linear actually tops the chart with a 66.48% success rate, edging out AlphaFold 3's 64.90%. AlphaFold 3 still keeps its crown on protein-DNA interactions (79.18% vs. 72.60%).

Alongside the folding system, the team also previewed SeedProteo, an all-atom diffusion model for de novo binder design based on similar architecture, outperforming BindCraft and RFDiffusion baselines across ten targets.

The work is detailed on their project site and arXiv preprint (2512.24354).

Why it matters

SeedFold makes a convincing case that classic scaling laws apply directly to molecular biology: make the representation wider and feed it massive distilled data, and precision climbs.

We think the linear attention work is particularly smart because reducing O(n³) complexity to O(n²) is essential if computational biology ever wants to scale to massive multimeric complexes without astronomical GPU bills. Computational biologists and drug discovery labs should keep a close eye on this, though the actual workflow impact will depend on how quickly reproducible checkpoints and inference code roll out to the community.