64.9%
#1 of 8
AlphaFold 3's FoldBench accuracy (2024, 370M parameters), the highest of eight models; SeedFold (2025, 923M) scored 63.1%
What the chart shows
FoldBench tests whether a model correctly predicts how a small molecule, such as a drug candidate, binds to a target protein. AlphaFold 3, released in May 2024 with 370 million parameters, scored 64.9%, the highest of the eight models in the data.
SeedFold (December 2025, 923 million parameters) came closest at 63.1%, followed by RosettaFold-3 (57.3%). Five later models are larger than AlphaFold 3 without scoring higher, which the AI Index reads as a sign that training data, not size, is the bottleneck.
Key numbers
| Measure | Value |
|---|---|
| AlphaFold 3 (2024, 370M) | 64.90% |
| SeedFold (2025, 923M) | 63.10% |
| RosettaFold-3 (2025, 730M) | 57.28% |
| Boltz-1 (2024, 610M) | 55.04% |
| Boltz-2 (2025, 521M) | 53.90% |
| Protenix v0.7 (2025, 368M) | 50.70% |
How to read it
- Accuracy on FoldBench’s small-molecule binding tasks as compiled by RAISE Health; scores on other cofolding tasks can rank models differently.
- Parameter counts are in the CSV; Protenix v0.7 is listed at 368M here and 370M in the model-size data.
Source: RAISE Health, 2026, via the AI Index Report 2026 (Stanford HAI), figure 6.1.9: FoldBench: protein cofolding performance, 2024–25. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion