GroundAnything is a 4B-parameter model that predicts structured labels and quantized coordinates through blockwise diffusion. Tokens within each block are refined bidirectionally and in parallel, while causal dependencies between blocks preserve support for long outputs and KV-cache reuse.
Across 30 benchmarks, its autoregressive variant averaged 72.42%. The diffusion variant averaged 61.75%, compared with 53.32% for the fast MTP-based LocateAnything model, while running at more than twice the speed of its autoregressive counterpart. An optional self-speculative mode reached a 4.51× speedup with a 0.74 percentage-point COCO F1mIoU drop, showing that grounding outputs need not be generated strictly left to right.