Quantization — running a big model in a small memory
Quantization stores weights at lower precision (8-bit, 4-bit) so a large model fits on modest hardware, trading a measurable accuracy drop for a large memory and speed win.
Not yet scoped.
Why
The PoC runs one model at several bit-widths and measures the accuracy-versus-footprint curve, the number that decides whether local inference is viable.
How it works
Not yet built.
Where it lands in Jayverse
Number: before deploying any local model for the research pipeline, run the accuracy-vs-footprint curve at a few bit-widths (8/4-bit) on real hardware, and pick the smallest bit-width that keeps accuracy above the decision threshold.
Devnet: if Devnet ever hosts inference alongside Anvil, budget hardware from the measured footprint curve, not vendor claims, so "runs locally" is verified rather than assumed.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
trade X for Y
X를 대가로 치르고 Y를 얻다, 맞바꾸다 · 정확도를 조금 내주고 다른 이득을 취할 때. "trading a measurable accuracy drop for a large memory and speed win"
footprint
(메모리·자원의) 차지하는 크기 · 모델이 실제로 얼마만큼의 자원을 쓰는지. "the accuracy-versus-footprint curve"
viable
실행 가능한, 현실적으로 쓸 만한 · 로컬 환경에서 돌릴 수 있는지 판단할 때. "decides whether local inference is viable"
modest hardware
사양이 평범한(고급이 아닌) 하드웨어 · 값비싼 장비 없이도 돌아간다는 뜻. "so a large model fits on modest hardware"