Compare

INT8 vs INT4 Quantization

Both shrink a model's weights from 16-bit floats to lower-precision integers to cut memory and speed inference — INT8 keeps more precision, INT4 keeps less and shrinks the model further.

DimensionINT8 QuantizationINT4 Quantization
Memory footprint

Roughly half the size of FP16 — a 7-billion-parameter model drops from about 14 GB to about 7 GB.

Wins here

Roughly a quarter the size of FP16 — that same 7-billion-parameter model drops to about 3.5 GB.

Quality preservation Wins here

Calibration is close to free in practice, typically under a couple of points of benchmark degradation.

Needs more careful calibration (methods like GPTQ or AWQ) and usually gives up more quality than INT8.

Hardware accessibility

Fits many models onto a single high-end GPU where FP16 needed several.

Wins here

Can put a model within reach of consumer GPUs, or a laptop, where INT8 still would not fit.

Inference throughput

Faster than FP16, though gains depend on hardware kernel support for the format.

Wins here

Generally faster still on memory-bandwidth-bound serving, though not guaranteed to double again over INT8.

Calibration complexity and tooling maturity Wins here

A simpler, more broadly supported quantization path with fewer failure modes.

Needs a more deliberate calibration process to keep quality loss manageable at this precision.

Risk on sensitive or edge-case tasks Wins here

Smaller precision loss makes it the safer default when you cannot extensively re-verify quality yourself.

Larger precision loss means edge cases are more likely to be where quality actually degrades.

Typical production choice

The right call when quality headroom is thin and hardware is not the binding constraint.

The right call when memory or hardware access is the binding constraint and you have validated quality against your own eval set.

When to choose INT8 Quantization

Choose INT8 when quality matters most, you have the hardware for it, and you haven't run your own quality eval to validate a more aggressive quantization.

When to choose INT4 Quantization

Choose INT4 when memory or hardware access is the binding constraint — self-hosting on limited GPUs, edge deployment — and you've validated quality on your own eval set.

The verdict

Quantize to INT8 almost by default — the quality cost is minor and the memory savings are real. Go to INT4 deliberately, and only after measuring your own model's quality delta on a real eval set, since the additional loss is real even where it is often acceptable.

Reading the trade-offs is step one.

Miatz turns comparisons like this into judgment — the free founding cohort trains you to make this call on real systems, not just recite it.