Cherimoya: Fast and Stable Small Model

jmschreiber91 · x · 2026-07-15

The Cherimoya model is incredibly compact: using 128 filters, it has fewer than 1 million parameters and can process over 40,000 samples per second.

The author highlights its implications for modern large-scale tasks (e.g., design): training takes only about 5–20 minutes, with very low cross-run variance and high stability.

Further details mention that Cherimoya is built on the Cheri block, a version of the ConvNeXT architecture adapted for genomic data. To further accelerate processing, the project provides two custom GPU kernels: one for training and a megakernel for inference.

Related event: Cherimoya: A Small, Fast Genomic S2F Model Reaching SOTA(6 posts)→

Original post →

More from Infra

Infra channel →