Cube is Roblox's open-source effort to build a foundation model for 3D intelligence — the 3D counterpart to the large models that now generate text, images, audio, and video. The ambition stated in the repository is broad: a model family that can help developers produce every part of an interactive experience, from generating 3D objects and scenes to rigging characters for animation to writing the scripts that describe how objects behave. Rather than ship a single closed system, Roblox is open-sourcing the models incrementally and inviting the research community to build toward that goal with it. ## Why a 3D Foundation Model Is Hard Text and images benefit from oceans of clean, web-scale training data; high-quality 3D assets are comparatively scarce and inconsistently formatted. A general 3D model also has to represent shape, structure, and semantics in a way that downstream engines can actually use — a generated mesh is only useful if it can be animated, simulated, and scripted inside a game. Cube's design goal is to produce assets that drop into a real engine and respond to animation, physics, and behavior code without heavy manual cleanup, which is a stricter bar than producing a pretty render. ## Cube 3D v0.5: Text-to-Shape The first released model, Cube 3D v0.5, focuses on shape generation from text prompts. It tokenizes 3D shapes into a discrete representation the model can predict, letting it generate a mesh from a natural-language description. Roblox ships the weights on Hugging Face along with an interactive demo, a Colab notebook, and a companion technical report, so the model is straightforward to try before integrating. As a v0.5 release it is explicitly an early checkpoint on a longer roadmap rather than a finished product. ## May 2026: CubePart The most recent addition, CubePart, moves from whole objects to controllable structure. It is an open-vocabulary, part-controllable 3D generator: given an input mesh and a user-defined parts schema, it synthesizes one mesh per schema element such that the pieces assemble into a coherent object while respecting the requested semantic breakdown. Because the output is already decomposed into meaningful parts, the resulting assets can be integrated into game engines and driven by animation, physics, and behavior scripts without manual post-processing — the exact engine-ready property that makes 3D generation useful for real development. ## Openness and Ecosystem Cube is developed in the open with code on GitHub, weights and demos on Hugging Face, and arXiv reports documenting each release. That transparency is part of the point: Roblox frames the project as an invitation to the research community rather than a product drop, and the incremental releases (v0.5, then CubePart) show an active roadmap. For a platform whose entire economy runs on user-generated 3D content, lowering the effort of creating that content is a direct strategic bet. ## Trade-offs and Limitations The honest framing is that this is early-stage research infrastructure. v0.5 versioning signals that shape quality, resolution, and controllability are still maturing, and the model does not yet cover the full "objects to scenes to rigging to scripts" vision the README lays out. Licensing is a custom Roblox research-oriented license rather than a permissive MIT/Apache grant, so teams must check the terms before commercial use. And like other 3D generators it targets GPU inference and assumes some familiarity with 3D asset pipelines. ## Who Should Use This Cube is most compelling for Roblox and game developers who want engine-ready generated assets, and for researchers working on 3D foundation models, shape tokenization, or part-level controllable generation. Anyone expecting a polished, production-grade text-to-3D product should treat it instead as a fast-moving open research platform — valuable for experimentation and contribution, still evolving toward the broader 3D-intelligence goal it describes.