WorldClaw is an AI agent system from Tencent Hunyuan that builds open, explorable and editable 3D worlds from a single written prompt. Every scene comes out as separate textured meshes ready for Blender, Unity or Unreal, rather than a video or a cloud of splats. For now it remains a research project, with no code, no demo and no pricing published. Game studios and 3D tinkerers can already browse the project page and the paper on arXiv.
WorldClaw splits world-building into three stages, planning, global terrain, then objects placed region by region. You type something like "a fishing village under snow", and planning agents turn it into a structured brief covering zones, terrain shapes, materials and spatial relations.
For detail-heavy areas, the pipeline renders the terrain, composes a reference image of what should sit there, cuts out each element, lifts it to 3D and drops it at the right spot on the ground. And when a cabin ends up hovering mid-air, a render-and-inspect loop catches it and fixes placement and proportions before handing the scene back.
Under the hood, the version described in the paper orchestrates Claude Opus 4.8 alongside GPT-Image-2, SAM3, SAM3D and Hunyuan3D, with everything assembled in Blender.
WorldClaw outputs independent 3D meshes you can export and edit one by one, unlike Gaussian splatting scenes or playable video worlds such as Google's Genie 3, which keep no recoverable geometry. Tencent's paper benchmarks itself against SynCity, Meta's WorldGen, MajutsuCity and World Labs' Marble, four systems already active in generated 3D worlds.
For a studio, the appeal is immediate. Prototype an entire map in plain language, then rework it by hand. The approach sits close to what 3D simulation tools have been chasing, with one big difference, here every tree and building remains a separate asset.
| System | Output type | Status |
|---|---|---|
| WorldClaw (Tencent) | Separate textured 3D meshes | Published research, no code available |
| Marble (World Labs) | 3D scenes generated online | Live commercial product |
| Genie 3 (Google DeepMind) | Playable real-time video world | No exportable geometry |
| HunyuanWorld 1.0 (Tencent) | 3D worlds from text or image | Open source since 2025 |
No code, weights, demo or API shipped with the WorldClaw paper on arXiv. The project page shows worlds in orbit view and ground-level walkthroughs, but there is nothing to download and no price list to compare.
Hope for an open release is reasonable, since Tencent already open-sourced HunyuanWorld 1.0. One practical detail worth knowing (the kind that cools down impatient tinkerers), even with the code in hand, running the published pipeline would require paid access to third-party models, and the paper's experiments ran on a server with four NVIDIA H20 GPUs.
Not yet. WorldClaw exists only as a research paper and a demonstration page, with no code, no weights, no online demo and no API, so no pricing has been announced either. Since Tencent has open-sourced other 3D projects such as HunyuanWorld 1.0, a code release remains plausible, though nothing official has been committed.
Marble is a live commercial product with a free tier and paid plans starting at $20 per month, while WorldClaw remains research. Tencent's paper highlights its per-object editable meshes as the point where it claims to go further than its rival's generated scenes, which it says lack independent object control.
No, despite what some YouTube coverage suggests. The paper itself benchmarks WorldClaw against four earlier systems that already generated editable 3D worlds, SynCity, WorldGen, MajutsuCity and Marble. Its real contribution lies in agent coordination and the render-based verification loop that corrects object scale and placement.
WorldClaw produces meshes and their placements, not game logic. Nothing in the pipeline generates scripts, gameplay or engine-specific collisions. The realistic workflow is exporting the scene to Unity or Unreal, then adding behavior by hand or with a coding agent on top of the generated world.
Verdict: Nothing to install today, plenty to watch tomorrow. Game studios, generative 3D researchers and Blender regulars should bookmark the project page while waiting for a possible code release.
