Roblox has recently launched the first version of its foundational AI model, Cube 3D, with plans to make it open source. This model aims to enhance the efficiency of 3D creation.

Cube 3D is trained to generate 3D objects. Its training draws inspiration from cutting-edge text generation models, which form sentences by predicting the next character (or set of characters). Similarly, Cube 3D tokenizes 3D objects, interprets shapes as tokens, and predicts the next shape token to construct complete 3D objects.

The training data for this model includes licensed datasets, publicly available datasets, and experiential data from the Roblox ecosystem.

In the future, Cube 3D will also be capable of generating objects from images. Roblox stated that this model will eventually evolve into a multimodal model, able to process various types of inputs such as text, images, and videos, and will integrate with existing AI creation tools.