I ported Hunyuan3D Shape/Paint to Mac, Swift-MLX, 1GB RAM in Q4

Author: ZimengXiongCreated Jul 11, 2026Updated Jul 27, 2026

benchmarks on M4 Max, verified parity with CUDA/pytorch-mps, these are FP16 sizes

run wall time peak memory
hy3d shape (small) 20.9 s ~5.6 gb
hy3d shape (large) 22.3 s ~7.3 gb
hy3d paint (rgb) 231 s ~38 gb
hy3d paint (pbr) 344 s ~39 gb

https://github.com/ZimengXiong/Hunyuan3D-Swift https://github.com/ZimengXiong/Modelr (app to run Hunyuan3D-Swift generation)

also on huggingface: https://huggingface.co/zimengxiong/hunyuan3d-mlx-paint-large https://huggingface.co/zimengxiong/hunyuan3d-mlx-shape-small https://huggingface.co/zimengxiong/hunyuan3d-mlx-paint-small https://huggingface.co/zimengxiong/hunyuan3d-mlx-shape-large

would love to integrate this somehow, but i feel it diverges too much and isn't even in python now

Source: Tencent-Hunyuan/Hunyuan3D-2