Explore AI

What Q4 actually costs you

Compression is the reason local AI exists. It is also the reason your local model feels dumber than the benchmarks promised.

THE SAME 8B MODEL, SIX WAYS: FP16 — the reference 16 GB, Q8 8.5 GB, Q5 5.7 GB, Q4_K_M — the compromise 5.5 GB, Q3 — quality starts to go 3.5 GB, Q2 — do not bother 2.8 GB

A model is a very large pile of numbers. Quantisation stores those numbers with fewer bits each, which is the only reason any of this fits on hardware you own.

The trade is simple to state and widely misunderstood in practice.

The scale

FormatBytes per parameter8B modelWhat it gives up
FP162~16 GBnothing — this is the reference
Q8~1~8.5 GBalmost nothing measurable
Q5~0.7~5.7 GBvery little
Q4~0.55~5.5 GBthe standard compromise
Q3~0.43~3.5 GBnoticeable, especially on reasoning
Q2~0.35~2.8 GBquality collapses

Q4 is the default for a reason. Above it you are spending memory for very little gain; below it you are spending quality for very little saving. The specific variant matters too — `Q4_K_M` is the one usually worth choosing.

Two things benchmarks will not tell you

Degradation is not uniform. Quantisation does not make a model uniformly slightly worse. It tends to hurt reasoning and instruction-following far more than it hurts fluency. A Q3 model can still write a fluent paragraph while failing to follow a three-step instruction — which is why it is easy to be fooled.

Published scores come from full-precision models. When a model card claims a benchmark result, it was almost certainly measured at FP16 or BF16. Your Q4 copy will score lower on exactly the tasks you care about, and there is no way to know by how much without testing it yourself.

The KV cache, which nobody mentions

The weights are only half the memory story. The KV cache is the model's working memory of your conversation, and it grows with context length. On a long document, the cache can exceed the weights.

This is why a model that "fits" on your hardware struggles with a long PDF: the weights fit, and the conversation does not.

Practical levers: cap the context length, and prefer models designed for long context rather than forcing it.

What to do about it

1. Start at Q4_K_M. Only go lower if the model genuinely will not fit. 2. If quality seems poor, try Q5 or Q6 before blaming the model. The difference is often the quantisation, not the weights. 3. Test on your own task. A model that scores well on a leaderboard and badly on your work is not a good model for you.

Where to go next