Why it matters
INT4 is what makes 70B models runnable on 24GB consumer GPUs. Understanding it opens up local LLM deployment.
Advertisement
The architecture
Storage: 2 weights per byte. Scale factor per group of weights.
Group size: 32-128 weights per scale typical. Smaller groups = better quality but more overhead.
Advertisement
How it works end to end
Naive INT4: quantize each weight independently. Poor quality.
GPTQ: calibration data + layer-by-layer optimization. Choose scales to minimize error.
AWQ (Activation-aware Weight Quantization): identify important weights, protect them. Even better quality.
GGUF: llama.cpp format, mixed precision (some layers INT4, others INT8, some higher).