NVFP4 decode kernel reportedly speeds generation by up to 10% while using more bits
@humansand claims better accuracy and says the kernel is already used in Ampersand’s RL and Boba serving.
TLDR
On Oct. 23, 2026, @humansand announced that it had accelerated an NVFP4 decode kernel while improving its accuracy. The account claimed up to 10% faster generation despite using more bits, a reversal of the typical trade-off it described. It said the kernel was already used in Ampersand’s RL and Boba serving.
Combined views
12.7K
10 Sources, first seen ago


