CoAtNet-4 backbone (Dai et al., 2021).

Even deeper variant: blocks_per_stage=(2, 12, 28, 2), dims=(192, 384, 768, 1536), stem_width=192, attn_heads=(24, 48). Approximately 275M parameters, 85.0% ImageNet-1k top-1 at 224×224 (Table 5). Paper headline with JFT-3B pretrain reaches 88.4% at 512x512.