A May 9, 2026 publication reports training a 15-million-parameter JEPA from raw pixels on one GPU and compares its planning speed on control tasks.
On May 9, 2026, a publication described LeWorldModel, a 15-million-parameter JEPA trained end to end from raw pixels. It reports using two loss functions and says the method plans up to 48 times faster than world models based on foundation models on control tasks.
The reported configuration and training approach may interest engineers developing compact world models. These figures and conclusions are the publication’s claims, not an independent verification; consult the original material and check its experimental conditions and comparisons before applying the results.