EPFL Introduces MODUS — A Universal Multimodal Model

Researchers from EPFL have developed MODUS, a decoder-only architecture capable of processing any combination of data through a dual-expert system.

Compiled by Sergey KostenchukPublished 2026-07-29Updated 2026-07-29

2026-07-29 Research HuggingFace

🤖 EPFL Introduces MODUS — A Universal Multimodal Model

Built upon BAGEL-7B, the decoder-only architecture utilizes two experts: a 1D expert for discrete tokens and a 2D expert for spatial data. The model supports 16 modalities and allows for the construction of transformation chains.

🌍 MODUS simplifies AI architecture by replacing complex systems with multiple encoders with a single unified decoder.

👤 This makes AI more flexible: for example, one can transform text into a depth map, and then use that map to create a realistic image.

Source 1: https://modus-multimodal.epfl.ch/ Source 2: https://github.com/EPFL-VILAB/Modus

Sources