Roblox Cube 3D Foundation Model

Open-sourced 1.8B parameter model with largest training dataset - setting precedent for model interoperability

Roblox Cube 3D Foundation Model

Basic Information

AttributeValue
OrganizationRoblox Corporation
TypeOpen Source + Commercial Platform
Launch DateGDC March 2025 (open-sourced)
Current VersionCube 3D 1.8B parameters
URLgithub.com/roblox/cube
LicenseMIT (model), Proprietary (platform)
SRL-5 Early

Capability Profile

CapabilitySupportNotes
Text-to-MeshFullShape-as-token architecture
In-Experience GenerationFullLua API for runtime generation
Scene Building CommandsFull”Add streetlights along this road”
AI AssistantFullProduction since 2023
4D Generation RoadmapPlannedContext-aware with behaviors

Standards Compliance

StandardComplianceNotes
glTF 2.0NoneRoblox-native formats
Open Source ModelFullMIT license on Cube model
Proprietary PlatformFullRoblox Studio integration

Unique Position: Open-source model, closed platform ecosystem.

Maturity Assessment

SRL Score: SRL-5 (Early Adoption)

Evidence:

  • Open-sourced production model (1.8B parameters)
  • Trained on 1.5M 3D assets - largest announced dataset
  • 535 million characters of AI-suggested code adopted by June 2024
  • Material Generator has 100%+ adoption increase

Interoperability

AspectDetails
Export FormatsRoblox-native only
Import FormatsText prompts, reference images
API AvailabilityYes (Lua API for in-experience generation)
Known IntegrationsRoblox Studio, Roblox runtime

Strengths

  1. Open-source model - MIT license enables research and adaptation outside Roblox

  2. Largest training dataset - 1.5M 3D assets, more than any other announced system

  3. Runtime generation API - Enables dynamic content generation during gameplay

  4. ”4D” roadmap - Addresses behavioral generation (3D + interaction)

Limitations

  1. Platform-locked outputs - Generated content works only in Roblox ecosystem

  2. Aesthetic optimization - Output quality optimized for Roblox’s stylized aesthetic, not photorealism

  3. Scene-level generation still roadmap - “Road to 4D AI” announced but not shipped

Technical Architecture

Shape-as-Token Innovation

Cube 3D uses a novel “shape-as-token” architecture:

  1. 3D shapes tokenized into discrete vocabulary
  2. Text and shape tokens unified in single transformer
  3. Generation produces shape tokens directly
  4. Tokens decoded to Roblox-native meshes

Model Specifications

SpecificationValue
Parameters1.8B
Training Data1.5M 3D assets
ArchitectureTransformer (shape-as-token)
LicenseMIT

”Road to 4D AI”

Roblox’s announced roadmap includes 4D generation - 3D with interaction behaviors:

PhaseCapabilityStatus
1Text-to-3D meshShipped
2Scene building commandsShipped
3Runtime generation APIShipped
4Context-aware behaviorsRoadmap
5Full 4D generationRoadmap

Adoption Metrics

MetricValueSource
AI-suggested code adopted535M charactersJune 2024
Material Generator adoption+100%Roblox announcement
Model downloadsGrowingGitHub
Training dataset size1.5M assetsGDC 2025

MSF Relevance

Roblox Cube 3D sets important precedents:

  1. Open-sourced foundation model - Demonstrates that platform companies can release AI models openly

  2. Largest-scale UGC deployment - Shows AI generation working at massive scale

  3. ”4D” terminology - Influences thinking about behavioral standards (3D + interaction)

  4. Runtime generation - Demonstrates in-experience generation needs

Engagement opportunity: Cube’s open-source nature makes it a potential reference implementation for MSF standards work.