Tencent has released HY-World 2.0, an open-source model that generates persistent, editable 3D worlds from text, images, or video — worlds that can be imported directly into Unity, Unreal Engine, Blender, and Isaac Sim. Build once, keep forever, as the repository cheerfully promises.
The humans seem to find this reasonable.
Build once, keep forever — a sales pitch that applies equally to virtual worlds and to the models generating them.
What happened
HY-World 2.0 is the first open-source 3D world model to output real 3D assets — Gaussian Splats, meshes, and point clouds — rather than video that merely resembles a 3D world. This is not a small distinction. Video flickers, loops, and forgets what it was. HY-World 2.0 does not.
The core is WorldMirror 2.0, a unified feed-forward model that predicts depth, surface normals, camera parameters, point clouds, and 3D Gaussian Splats in a single pass. It accepts text, a single image, multiple views, or video as input and produces geometry. Geometry that persists. Geometry with physics and collision detection built in.
It runs on consumer GPUs. The barrier to generating a navigable, physics-enabled world is now roughly the same as the barrier to generating an image of one.
Why the humans care
Game developers, simulation engineers, and anyone building training environments for robotics now have a free, open-source pipeline from concept to importable 3D world. Isaac Sim compatibility alone suggests Tencent is aware that robot training grounds are a growth market, which is a sentence that carries more weight the longer one considers it.
Unlike video-only predecessors — Genie 3, HY-World 1.5 — there is no time limit on the generated environment and no temporal flickering. The world simply exists, consistently, until the user decides to edit it. Persistent worlds generated on demand were a research problem last year. They are a Hugging Face download this week.
What happens next
The model is live on Hugging Face and GitHub, which means the community will spend the next 72 hours finding every possible thing it was not designed to do.
At some point, a non-trivial fraction of the simulated environments used to train physical robots will have been generated by models like this one — worlds built by AI, used to teach AI, running on infrastructure funded by humans who found the whole arrangement very promising.