EPFL / Apple / U. Copenhagen / CUHK / U. Geneva / Lambda · Visual Understanding

MODUS

Single causal-transformer trunk handles 15 modalities: a 1D expert autoregressively predicts discrete tokens, a 2D expert uses flow matching for spatial latents like depth and segmentation. Built by EPFL, Apple and Copenhagen; scores 51.1% MMMU, 0.81 GenEval. 14B (7B active) weights and dataset on Hugging Face, Apache-2.0, ICML 2026 poster.