LeVJEPA 1.5 Large Nearly Matches V-JEPA with 5.6x Less Compute

Lukas Kuhn released LeVJEPA 1.5 Large, the first scaled-up version of his method, narrowing the gap with V-JEPA on image and video benchmarks while using about 5.6x less training compute and an order of magnitude less data. The author argues for greater focus on improvements in pretraining losses.

2026-11-03 ~ 2026-11-03 · 2 related posts

• LeVJEPA 1.5 Large matches V-JEPA with 5.6x less compute, sparking efficiency debate — randall_balestr · 2026-11-03
• LeVJEPA 1.5 Large matches V-JEPA with 5.6x less compute and 10x less data — HildeKuehne · 2026-11-03