Why the official code is CUDA-only, what a Mac port takes, and the shortcut.
TRELLIS.2 is Microsoft's 4-billion-parameter image-to-3D model and one of the best all-round engines available. The catch on a Mac is that its reference implementation is built around NVIDIA CUDA: custom sparse-voxel and rasterization kernels that have no Apple Silicon equivalent. Running the official code on a Mac is not a matter of a flag, it needs a real port. Here is what that involves, and the shortcut.
The reference repo depends on CUDA-only components: custom sparse-structure and octree rasterization kernels, fused attention, and more. These are compiled for NVIDIA GPUs and do not exist on Apple Silicon out of the box.
To run it locally without an NVIDIA GPU you either rent a cloud GPU (which defeats the point of local), or port the model to a Mac-native backend such as Apple's MLX or PyTorch MPS, reimplementing those CUDA kernels. That is a serious engineering project, not an install.
A Mac-native run needs the weights in a form the new backend can load, plus a matching sampler and mesh-extraction path. Reaching parity with the reference output is the hard part.
Unlike smaller models, TRELLIS.2 will not run on a Mac by swapping one dependency. Its CUDA kernels are the model. This is exactly the gap 3D Mate closes.
3D Mate ships a Mac-native TRELLIS.2 4B built on Apple's MLX, so the full 4B model runs on Apple Silicon with no CUDA, no port, and no cloud. Install it from inside the app and generate offline; a balanced result takes about 8 to 10 minutes on an M1 Pro.