Aug 5, 2026Research
Tencent's open-source world model speaks splat natively
Illustration: Splats Newsroom · Research
Tencent Hunyuan updated HY-World to 2.1 — the latest version of its open-source world model whose output is editable 3D, meshes and gaussian splats, rather than generated video.
Why it matters: Most of the world-model race produces video frames you can't touch. Tencent is the largest player shipping open-source world models whose native output is 3D assets — a direct bridge between the world-model boom and the gaussian splatting ecosystem.
Zoom in:
- World generation runs through a four-stage pipeline: panorama generation, trajectory planning, world expansion, and composition — from text, images, videos, or multi-view photos.
- WorldMirror 2.0 handles reconstruction, predicting dense point clouds, depth maps, surface normals, camera parameters, and 3DGS in a single forward pass.
- Reported model sizes: HY-Pano-2 at roughly 80B parameters, WorldStereo-2 at ~17B, and WorldMirror-2 at ~1.2B.
Catch up quick: Tencent open-sourced the stack in stages this year: the technical report, partial code, and WorldMirror 2.0 weights landed April 16, HY-Pano 2.0 weights on May 11, and world-generation code with WorldStereo 2.0 weights on May 18.
Go deeper:
