4M-21

4M-21 is an any-to-any multimodal foundation model trained across 21 modalities and tasks (RGB, depth, surface normals, segmentation, edges, captions, bounding boxes, DINO/CLIP features, metadata and more) via masked modeling over discrete tokens. It is the reference open any-to-any model — a single network that generates or conditions on any subset of modalities. From Amir Zamir's VILAB (joint→Apple); NeurIPS 2024. The VILAB/Apple any-to-any line continues with FlexTok, which is tracked under Apple.