ByteDance Unveils Next-Gen Folding Model SeedFold, Surpassing Google's AlphaFold 3 in Performance
2026-01-17 / Read about 0 minute
Author:小编   

ByteDance's Seed research team has unveiled SeedFold, a cutting-edge, next-generation folding model that introduces a groundbreaking methodology for scaling foundational models within the realm of biomolecular structure prediction. This innovative model has demonstrated superior performance compared to AlphaFold 3 across a variety of protein-related tasks. The team has identified three pivotal 'scaling strategies':

  • Model Scaling: Emphasizes breadth over depth, with empirical evidence indicating that widening the model's backbone yields more favorable outcomes.
  • Architectural Innovation: Incorporates a linear triangular attention mechanism, effectively reducing computational complexity from a cubic to a quadratic level.
  • Data Expansion: Leverages extensive knowledge distillation techniques to expand the training dataset to an impressive 26.5 million samples.
On the FoldBench benchmark test set, SeedFold and its linear variant showcased top-tier performance across several critical tasks. Researchers have highlighted two promising avenues for future exploration: mixture-of-experts networks and post-training scaling. SeedFold sets a solid foundation for the development of the next generation of biomolecular foundational models.