LeVJEPA Cuts Video Pretraining Compute by Up to 20.8x, Matching V-JEPA 2

Yann LeCun's team introduced LeVJEPA, the first video encoder trained with the LeJEPA collapse-free objective, eliminating EMA target encoders and stop-gradient tricks. It cuts pretraining compute by 5.6x–20.8x while matching V-JEPA 2; first author Lukas Kuhn will present at Cohere Labs on September 8.

2026-09-03 ~ 2026-09-03 · 3 related posts

• LeVJEPA: video encoder matches V-JEPA 2 with 5.6-20.8x less pretraining compute — CSProfKGD · 2026-09-03
• LeVJEPA: collapse-free video pretraining matches V-JEPA 2 with 5.6–20.8x less compute — Cohere_Labs · 2026-09-03
• LeVJEPA from LeCun's team: video pretraining matches V-JEPA 2 with up to 20.8x less compute — Cohere_Labs · 2026-09-03