building the strongest world model through generative game worlds

Iteration throughput is the bottleneck for world models

Physical robots are too slow

Physical feedback is expensive and sparse.

Lab-built worlds are too narrow

In-house scenarios are costly and limited.

Gaming throughput =

Diversity × Correction × Scale

01

Scenario diversity

Long-tail worlds created by players.

02

Correction speed

Failures fixed during play.

03

Scale

Entertainment-scale usage, not paid data work.

Today's game-making paths are broken

Path 01

Traditional engines

Expert-only creation Misses physical reality

Path 02

Pixel video models

No consistency Weak control Impossible cloud economics

Our thesis

The strongest world model requires a new foundation model architecture, evolved through generative gaming.

JST: better world compression through tokenized 4D atoms Our novel proprietary Joint Space & Time foundation model

Generative foundation model

Learns realism and style from data, without hand-built graphics pipelines.

Physics-native

Models atoms instead of pixels for guaranteed consistency and controllability.

On-device inference

Runs locally on iPhones and personal computers.

JST solves one bottleneck per generation

Graphics, physics, and reasoning mature together in the same model.

JST-1 Q4 2022 - Q2 2024

Solve architecture

Character-only proof of concept for joint space-time modeling.

JST-2 Q2 2024 - Q2 2026

Solve graphics

Characters, scenes, motion, and asset rendering with game-ready 3D physics simulation and planning.

JST-3 Q2 2026 - Q4 2027

Solve physics

Generative physics engine for interactive real-world simulation beyond traditional engines.

JST-4 Q4 2027 - Q1 2029

Solve reasoning

General physical intelligence grounded in graphics, physics, and interaction.

Graphics Proof Solved Solved Solved

Physics Basic Game-ready Solved Solved

Reasoning Basic Early Growing Solved

JST-1 proved the architecture in the real world

A character-only JST model became a profitable, self-sustaining global AI application.

45M+ users

250M+ videos generated

$2M ARR

$1.2M total annual cost, including all free users

Architecture proof

Character-only model with consistency, controllability, and efficiency.

Global validation

Meme Maker reached massive usage with zero paid acquisition.

Economic proof

Self-sustaining after serving the entire free viral funnel.

JST-2 demo

JST-2 is the strongest playable world model today

vs Marble ($5B val.)

Motion + behavior

MarbleStatic scenes.

JST-2Scenes + dynamic characters with text-driven behavior.

vs Marble ($5B val.)

Feed-forward generation

MarbleVideo model + optimization; >5 min and >$1 per scene.

JST-2Feed-forward foundation model; hundreds of ms and <$0.01 per scene.

vs Decart ($4B val.)

On-device inference

DecartCloud GPU inference + video streaming.

JST-2Runs on phones with zero marginal cost and zero streaming latency.

The team to build playable world models

Hang Chu, CEO

Leading researcher in 3D generative models for 10+ years.

Former Principal Researcher, Autodesk.
Google, NVIDIA, Facebook.
Cornell MS; PhD researcher, University of Toronto ML Group (5 years).
2.6K+ citations; h-index 19.

Ming Liang, CTO

Researcher and systems builder for 3D foundation models.

Former Staff Scientist, Uber ATG.
Waabi, Apple SPG.
Tsinghua PhD; Kaggle Data Science Bowl champion ($1M prize).
11.6K+ citations; h-index 31.

Core team

**Jinma** Data Lead 10+ years as ML engineer

QQ Model Lead 3.8K+ citations; 10+ yrs research

**Yun** Chief Scientist First author, CVPR Best Paper finalist

Sharp Runtime Lead Former NVIDIA CUDA team

**Yao** Infra Lead Former Sr. Engineer, ByteDance

**Jason** Product Lead Former AI startup founder

**Nan** Growth Lead USC + LSE dual master's

Advisors

**Eric** Strategy Former CTO, Ctrip; ex-eBay

**Christine** GTM First marketer at Twitter

Model-as-a-Service for generative worlds

Prosumer

Subscriptions

Paid plans for higher throughput, fidelity, and control.

B2B + developers

API access

Usage-based access to JST generation, animation, and simulation primitives.

Enterprise

Licensing

Embedded JST models for large creative, gaming, and interactive platforms.

2026 year-end target $10M ARR

2027 year-end target $100M ARR

The ask: $200M to train JST-3 and scale Model-as-a-Service

$80M 40%

JST-3 training compute

$40M 20%

Data + simulation infrastructure

$40M 20%

Research and engineering

$40M 20%

Product distribution + GTM