LLM Efficiency · Compression

Weight Quantization

Drop the bits — same weights, coarser grid. Watch size shrink and precision blur.

precision
weight value −1+1
levels (2^bits)
model size · 7B params
precision · step
quantize.py