Hunyuan3D 2.1

Two models in a pipeline. Hunyuan3D-Shape-v2-1 is a 3.3B diffusion transformer, 21 transformer layers with an MoE layer for latent representation, that turns a single image into a watertight mesh. Hunyuan3D-Paint-v2-1 is a 2B mesh-conditioned multi-view diffusion model that generates physically based rendering materials over that mesh, so what comes out has albedo, metallic and roughness channels rather than a single flat texture. Shape generation alone runs in 10GB of VRAM, texture generation in 21GB, the full pipeline in 29GB. One 32GB card covers the lot.

Tencent released the full training code with the weights in June 2025. That is the part almost every other open 3D release skips, and it means fine-tuning on a proprietary asset library is a normal engineering task rather than a research project. Hunyuan3D-Omni later extended the same backbone to accept point clouds, voxels, bounding boxes and skeletal poses as conditioning signals, which is how you get deterministic control over geometry and pose instead of rolling prompts until something usable appears.

Now the licence, because it decides whether any of the above is available to you. The Tencent Hunyuan 3D 2.1 Community License Agreement is not open source in the OSI sense. It defines a Territory that excludes the European Union, the United Kingdom and South Korea, and forbids using the model, its derivatives or its outputs outside that Territory. Cross 1 million monthly active users and you must request a separate licence from Tencent. Outputs may not be used to improve any other AI model. Note also what never opened: Hunyuan3D 2.5, the PolyGen quad-mesh model announced 8 July 2025 and the 10B Hunyuan3D 3.0 with 1536-cubed geometric resolution are all hosted product or API. The open branch has not moved since 18 June 2025.