4×
Int4 quantization
Weight size cut by four, little quality loss
7×
Mixture of Experts
70B total, 10B active per token
∞
Partial loading
Model streams from flash on demand
2–3×
Distillation
Small models trained on big-model outputs
3–5×
Sparsity
Zero out and skip unused weights at inference
stack
All of the above, together
Compound effect: ~10–50× effective capacity on the same silicon