China's Hi3D open-sources V1.0 3D model trained on about 2 million assets

On Oct 27, Hi3D released the code and weights for Hi3D V1.0, an image-to-3D model the team describes as one of the earliest completed entries in its lineup and one already validated inside production workflows. The release is on GitHub and Hugging Face, alongside an official project page where the code can be downloaded and used free of charge.

The technical lineage runs back to Sparc3D, a high-resolution 3D modelling framework published at NeurIPS 2025 that concentrated on representing and reconstructing complex geometry so that small parts and surface detail survive. V1.0 keeps that core architecture; the work went into data scale, high-resolution training and production validation instead.

The most concrete change is training data. Where the Sparc3D paper used 500,000 assets, Hi3D V1.0 trains on roughly two million curated 3D assets, about four times as many. That matters for single-image generation: one reference view only shows part of an object, and the model has to infer the back and occluded regions from learned shape priors, so more varied assets give those inferences better support.

Generation runs coarse to fine. An upstream model first produces a rough mesh that fixes the overall silhouette and main parts, and V1.0 then refines geometry sparsely around the surface while consulting both the reference image and that coarse mesh. It extracts a mesh at every level and climbs through 1024³ and 1536³ stages, where the numbers describe the resolution of the geometry extraction grid rather than the input image.

In its technical report, the team puts multi-view fidelity at 0.75 and detail richness at 0.841, ranking first among the methods listed in its comparison, with a 90% slicing success rate for preparing meshes for 3D printing. Slicing success only measures readiness for print preparation; whether a model actually prints still depends on wall thickness, supports and materials.

Hi3D also showed V3.0, which it calls the industry's first commercially usable 2048³ domestic AI 3D model, paired with Hi3D Agent, an assistant built for 3D creators. In QbitAI's side-by-side tests on the same reference images, the Agent and V3.0 outputs stayed closer to the original in outline, volume relationships and local detail, while the comparison run against "GPT-6 Astra" produced only a recognisable base form.

The agent covers intent parsing, task planning, generation, iterative edits, engineering optimisation and delivery. It can extend assets into playable game demos, lightweight and rig characters for animation, or carry a model through part separation, connection design, colouring and print preparation for physical figures, and generated results can be exported to Blender or a game engine for further work.

Beyond the open model, Hi3D exposes MCP connections so external agents and tools can submit generation jobs, poll progress and retrieve results, which fits its positioning as AI 3D infrastructure rather than a single app. The open question is adoption: whether researchers treat V1.0 as a reproducible baseline, and whether developers and enterprises actually build on the commercial licence.