#mxfp4

Crusoe Keeps 4-Bit Weights to Cut LoRA Memory by 75%

Crusoe shows how storing frozen base weights in 4-bit and dequantizing in-kernel cuts LoRA memory by up to 75% without accuracy loss.