Reka Labs lays out Omni-World Models: one architecture for language and physical AI

RekaAILabs · x · 2026-09-17

Reka Labs argues LLMs are merely an intermediate step toward physical-world intelligence and proposes "omni world models" — a single unified architecture handling language, image, video and actions as both inputs and outputs. Key challenge: jointly modeling discrete signals (text, autoregressive) and continuous signals (audio, pixels, diffusion). Unifying VLA and world-model approaches under one backbone, they contend, is necessary to capture motion and geometry in the real world.

Original post →

More from Multimodal

Multimodal channel →