LeWorldModel open-sources stable pixel world model

This is a meaningful research code drop, not a shiny product launch: the value is in making a fragile world-model idea reproducible without the usual bag of tricks.

• –End-to-end pixel training without pretrained encoders or EMA machinery lowers the barrier for labs and independent researchers to iterate on JEPA ideas
• –The reported scale is practical for experimentation: roughly 15M parameters on a single GPU in a few hours
• –Shipping checkpoints, configs, and evaluation scripts makes the release more useful than a paper abstract alone
• –The strongest signal here is architectural, not commercial: stable compact world models are still an open problem, and this repo shows one plausible path