Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

K-Culture GlossaryTechnical words in the news

Quantization

Compression that shrinks a model by lowering the precision of its numbers — the trick that lets LLMs run on phones.

In plain words

Quantization is a compression technique that reduces size and computation by lowering the precision of the numbers inside a model. Instead of storing weights as 16-digit decimals, you store them as 4- to 8-bit integers — similar to lowering a photo's quality a notch to shrink its file size drastically.

The key point is that performance barely drops even after cutting precision. The fact that LLMs can run on laptops and phones, and the community culture of running open models on personal PCs, all exist thanks to quantization. When you see a phrase like "4-bit quantized version released," read it as "now you can run this on your home computer."

Distillation vs. Quantization: What's the Difference?

Lowering the precision of an existing model, Training a new, smaller model

Both get described as 'making a model lighter,' so they look like the same technique. But one means training a separate, smaller model to learn from a larger one's abilities, while the other means shrinking the size of an existing model by coarsening its numbers. They're often combined too — distill down to a small model, then quantize it further.

AspectQuantizationDistillation
What it doesReduces the number of digits used to represent the model's numbersTrains a smaller model using a larger model as a teacher
The resultA lighter version of the same modelA new model, distinct from the original
What it costsJust a conversion — much fasterRequires training — time and hardware
What's lostPrecision is trimmed, slightly blurring answersFalls short of the teacher's ability
In articles"A 4-bit version that runs on phones""A small model distilled from the large model is released"

Rule of thumbIf a new model is created, it's distillation; if the same model becomes lighter, it's quantization.

See also

Stories using this term

No story has used this term yet. New ones attach here automatically.

Browse every entry