Hi3D V1.0 (formerly Sparc3D) goes open source with mesh refinement up to 1536³

What this is about

Earlier this week, the company Hi3D open-sourced its 3D model Hi3D V1.0. According to the press release, that happened on Oct. 24, 2026. The model was formerly called Sparc3D and drew attention in 2025 for its high-resolution 3D geometry. Source code and model weights (the numbers the model learned during training) are on GitHub and Hugging Face. The press release also says you can generate 3D models for free on the Hi3D website. It lists use cases such as 3D printing, e-commerce, industrial design, VR/AR/XR, animation and film.

The catch: the press release sounds like an all-purpose generator. But the technical analysis of the repository by the OrcaRouter blog shows a more specialized tool.

What the model actually does

Hi3D V1.0 is not an image-to-3D generator in the usual sense. It is a refinement stage. You supply a reference image of an object and a coarse mesh (a triangle net describing the surface) of the same object. You get a higher-resolution mesh back.

Technically, the image conditions a "sparse flow" that works on the coordinates of the coarse mesh. "Sparse" means only regions near the surface are computed, not the whole volume. The pipeline draws sparse latents (compressed internal representations), decodes them through a self-pruning sparse VAE, extracts the surface and re-voxelizes it (voxels are 3D pixels) for the next pass.

Each pass uses 25 scheduler steps by default. According to the README, 1536³ is a nominal coordinate and extraction grid, not a dense tensor and not a fixed triangle count. The output is a trimesh.Trimesh, centered and normalized to unit radius, so scale and world coordinates are lost.

Vendor claims and what is actually backed up

The press release makes four numerical claims:

• The training dataset grew from about 500,000 to two million curated 3D objects. This appears on the Model Card as a project statement.
• The geometry reaches an effective sparse grid of up to 1536³. The pipeline does accept this setting.
• 90% slicing success rate. This figure appears nowhere in the repository. There is no definition, no slicer and no baseline.
• "Highest scores for fidelity and detail richness" in the reported evaluation. There are no numbers, no comparison models and no methodology.

The repository contains only inference code, a demo, Docker constraints, docs and the Python package sparc3d. Training code, evaluation tools and benchmark data are missing. The vendor does not state a parameter count anywhere. So these claims are unconfirmed, not disproven. No independent verification exists yet.

The weights come in eight files (index, scheduler, VAE and transformer configs, model.safetensors, plus weights for transformer_512 and transformer_1024). Loading uses JSON and safetensors. It neither runs Python code from the model repository nor unpacks pickle checkpoints, which is better from a security standpoint. The conditioning encoders facebook/dinov2-large and openai/clip-vit-large-patch14 must be downloaded separately and have their own licenses.

Limitations according to the README

• Geometry only: no textures, materials or vertex colors; batch size one.
• A coarse mesh is required. The model does not invent an initial surface from an image.
• Errors in the coarse mesh persist and can become sharper.
• Areas hidden in the image are filled in plausibly, not reconstructed exactly, so it is unsuitable for measurements.
• Absolute scale is lost.
• The same seed does not guarantee the same mesh, because GPU operations are not bitwise deterministic.
• If the requested resolution can't run, you get an error instead of a silent downgrade.
• According to the README, resolution and detail do not guarantee suitability for your target application.

License

The "Hi3D Open License Agreement Version 1.0" is not an OSI-approved license. Commercial use is allowed. If a product that uses it exceeds 200,000 monthly active users or USD 1 million in monthly revenue, you must notify Hi3D and apply for a separate written license within 60 days. Commercial integrations must display and link "Powered by Hi3D 1.0". Pure outputs that contain no Hi3D material don't trigger these obligations. DINOv2, CLIP, Hunyuan and other components keep their own terms.

What you need to try it

• Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or higher, and CUDA Toolkit 12.0–12.5 matching PyTorch.
• Measured with Python 3.10.12 and PyTorch 2.5.0 (CUDA 12.4) on a 96 GB H20. Whether smaller cards, such as 24 GB, are enough is not documented.
• No CPU mode, no Mac, no Windows support mentioned. Google Colab with a large GPU is only conceivable if enough VRAM is available (not documented).
• Installation from source: flash-attn 2.7.4.post1 (build after PyTorch, no fallback), spconv-cu120 and a constraints file. No PyPI package.
• For the demo, a coarse mesh is optionally generated with Tencent's Hunyuan3D-2 mini (obtained separately). Your own meshes work too.
• Disk space: not stated. The weights plus DINOv2 and CLIP must be downloaded.

There is no hosted API, and on Hugging Face the model is not set up with any Inference Provider. Commercially, Hi3D sells the V3.0 model (2048³ voxels, textures up to 8K, online workspace). According to the announcement, that includes a 3D agent, MCP integration and a cut in credit consumption of over 50%.

Our take

The release is interesting for teams with a mesh pipeline and a big GPU, especially in 3D printing and asset production. For casual users, the bar is high. On the plus side: the README is honest about limitations, and the weights load safely. On the downside: the press release promises more than the tech shows, key figures (90%, "highest scores") are unsupported, the license is not OSI-compliant and has a growth clause, and the free version is the older one, while the current V3.0 remains paid. A solid verdict on quality has to wait until third parties test it.