Q4 vs Q8: How Much VRAM Do You Really Need?

Quantization is the single biggest lever for fitting a model on your GPU. In plain English: it shrinks the model so it uses less VRAM, with a small hit to quality. Q4 is the sweet spot for most people.

FP16full size Q8~53% Q4 ~28% · best value smaller = fits more GPUs

What quantization actually is

A model stores billions of numbers (weights). Quantization stores each number with less precision — think of it like saving a photo at slightly lower quality to make the file smaller. Q4 = 4-bit, Q8 = 8-bit. Fewer bits = smaller file = less VRAM.

How much VRAM each quant needs

Roughly, for a 13B model:

VRAM FOR A 13B MODEL

Q4_K_M
~8 GB
Q5_K_M
~9.5 GB
Q8_0
~14 GB
FP16
~26 GB

Add ~1–2 GB for context. Numbers scale with model size.

Which one should you use?

QuantQualityUse it when
Q4_K_M~98% of fullDefault. Best size-to-quality. Start here.
Q5_K_M~99%You have VRAM to spare and want a touch more.
Q8_0~99.5%Quality-critical work and plenty of VRAM.
Q3 / Q2Noticeably worseOnly to squeeze a bigger model onto a small card.
Bottom line: pick Q4_K_M unless you have a specific reason not to. The quality loss vs Q8 is tiny; the VRAM savings are huge.
Want the exact quant that fits your GPU + model? The calculator does the math.
Check your GPU free →

New to this? Start with How to Run an LLM Locally, or see what a 12GB GPU can run.