Verified, then logged.
Sun 22 Nov 2026 · 00:46 KST / 15:46 UTC한국어 | ENSearch /

Breaking01 / AI

Compressed translation models take up to 5.3 times as long

The study abstract reported lower memory use and changes of about one point or less in most aggregate translation scores.

Researchers reported on arXiv on 19 Nov 2026 that translation took 1.3 to 5.3 times as long after model compression in their tests.

The study abstract covered translation tests in Hausa, Swahili and Yoruba. The researchers used quantisation to lower the numerical precision of models after training. They compared INT8 and INT4 with an FP16 baseline on an NVIDIA T4 GPU.

The abstract reported that INT4 cut memory use by 49–61%. Aggregate chrF scores, which measure translation quality at the character level, generally changed by about one point or less. The largest change was 2.19 points.

According to the abstract, the dedicated translation model outscored both large language models in every language tested. Its lead in Hausa and Swahili ranged from 32.7 to 43.3 chrF points.

The researchers did not report results by type of translation error in the public abstract.