The release includes inference code in the hi3d-official/Hi3Dv1.0 repository and downloadable weights on Hugging Face.
The model takes a coarse mesh plus a reference image of the same object and refines the geometry. Hi3D says it progressively samples sparse latents, uses a self-pruning sparse VAE, and extracts meshes at sparse grid resolutions up to 1536 cubed across three passes. Its training data reportedly grew from roughly 500,000 to roughly 2 million curated 3D assets. The Python package retains the sparc3d and Sparc3DPipeline names.
The weights use a custom Hi3D Open License Agreement 1.0, rather than an OSI-approved licence. Training code and the evaluation harness are not included. The announcement also claims that Hi3D's commercial V3.0 reaches 2048 cubed voxels and that its hosted service reduced credit use by more than 50 percent; the source does not provide independent validation of those claims or benchmark results for v1.0.