Flagship Pioneering(@FlagshipPioneer):These figures show peak memory and runtime for a range of co-folding models, all on the same 96 GB GPU. What the grey curves have in common is more interesting than what separates them. They bend sharply upward past about 1,000 tokens and reach the card's ceiling between 2k–3k residues. After this threshold, these models stop working. Why? Because they share an architecture that spends memory as a square of the size of the complex they are attempting to fold. LightFold’s near-linear memory allows that same GPU to keep folding complexes past 15K residues on the same hardware, so its line rises gently. Taken together, where the other models hit a cliff, LightFold finds a slope that allows it to move beyond folding subunits into folding whole assemblies. More here:

2026.08.24 16:11

These figures show peak memory and runtime for a range of co-folding models, all on the same 96 GB GPU. What the grey curves have in common is more interesting than what separates them. They bend sharply upward past about 1,000 tokens and reach the card's ceiling between 2k–3k residues. After this threshold, these models stop working. Why? Because they share an architecture that spends memory as a square of the size of the complex they are attempting to fold. LightFold’s near-linear memory allows that same GPU to keep folding complexes past 15K residues on the same hardware, so its line rises gently. Taken together, where the other models hit a cliff, LightFold finds a slope that allows it to move beyond folding subunits into folding whole assemblies. More here: