weights.zip shrinks neural-network checkpoints by up to 34 percent, and by up to 87 percent on pruned or sparse models. Every compressed file restores bit-for-bit identical to the original, verified on the way back out, so what you load is exactly what you trained.

At scale the math is hard to ignore. Store 100 PB of checkpoints and you keep roughly 66 PB, trimming about 500,000 dollars a month off a typical hot-storage bill. The savings compound across every replica, every snapshot, and every version you retain.

Decompression runs at GPU speed, 45 milliseconds for a whole 27B model on an H100 (new decoder, next release), so compressed models load faster than uncompressed ones, even from host memory. Because compressed weights stay resident in GPU memory, a single card can serve a model roughly a third larger than its VRAM would otherwise allow.

It drops into the workflow you already have. Point it at your safetensors files, turn on verify mode, and ship. In head-to-head tests it beat zstd on ratio. Smaller bills. Faster loads. Zero loss.