Yesterday we announced Bonbon, and a lot of the conversation focused on the 61.5% hit rate. We understand why, but for us the other exciting result is the one in this chart.
We trained Bonbon on protein and molecule sequences alone, without any structures or labels. At smaller sizes, its scores on these tasks barely changed. When we scaled it to 1.6B parameters, they jumped, even though the training loss gave no sign that this was coming. It makes us think there is a lot more in these sequences still to find.
Adi will present Bonbon as an Oral at #NeurIPS2026 . If you missed the blog, you can find it in the replies.
We trained Bonbon on protein and molecule sequences alone, without any structures or labels. At smaller sizes, its scores on these tasks barely changed. When we scaled it to 1.6B parameters, they jumped, even though the training loss gave no sign that this was coming. It makes us think there is a lot more in these sequences still to find.
Adi will present Bonbon as an Oral at #NeurIPS2026 . If you missed the blog, you can find it in the replies.
0:10