Physics-native world foundation model

The first complete end-to-end model stack.

Physical AI has a quadrillion-year experience gap

Physical AI can only scale through interactive learning in a world simulator.

LLM World modelexperience engineJST-2, World Labs
video models, Meshy/Tripo

Physical AIrobot brainJEPA, VLA/WAM + hardware Embodied AI

10⁹ living beings in parallel

×

10⁶ years of evolution

=

10¹⁵ years of world experience

Biological intelligence was shaped by this. Physical AI has no scalable way to acquire it.

Real robot pre-training

Not a viable path at this scale.

Accurate + efficient world simulator

The only path that scales.

The real test of a world model is generality

The endgame is a full-stack physics-native model. Others add 3D only at the output.

Tokenization Architecture Data Engine Pre-training Post-training

JST-2

World Labs / SpAItial AI

Video models

physics-native partial none

Adjacent approaches still leave world simulation unsolved.

JEPA

A learning architecture, not a complete world simulator.

Tokenization, data, and end-to-end training for world simulation remain unsolved.

Meshy / Tripo

Built to generate individual mesh objects for traditional game engines.

Their generality stops at individual objects, not physical, dynamic worlds.

Embodied AI

Real-world learning is slow, expensive, and embodiment-specific.

As hardware evolves, data loses transferability; intelligence and embodiment must be iterated together inside a world simulator.

JST unlocks a new scaling law for world simulation Video models learn pixel correlations. Physics-native foundation models learn the world's underlying structure, delivering higher accuracy while reducing cost by orders of magnitude.

Model performance vs. cost

Pixel and reconstruction models share one cost-quality frontier. Physics-native foundation models open a new one by changing the representation, not the implementation.

Why existing routes cannot change the curve

Video models

Pixels in, pixels out. Without a persistent 3D representation, consistency and control cannot be guaranteed. Fidelity is probabilistic. More compute buys more attempts, not certainty.

World Labs / SpAItial AI

The base remains a video model, with 3D reconstructed at the output. Video generation plus reconstruction makes every scene roughly 1,000× slower than JST-2.

Embodied AI

Real-world pre-training is bottlenecked by physical time, hardware throughput, and embodiment-specific data. It cannot reach the scale or diversity physical intelligence requires.

JST-2 crossed the first generality threshold

Generality begins with one foundation model spanning characters, motion, and scenes.

JST-2 Meshy Tripo Cartwheel Hunyuan Marble SpAItial AI Character Image-to-character — — — — Rigging — — — — Motion Text-to-motion — — — — Video-to-motion — — — — — Scene Text-to-scene — — — — Image-to-scene — — — —

Only JST-2 covers all three primitives and all six foundational tasks.

One checkpoint, across-the-board SOTA

JST-2 beats every specialist benchmarked, with no task-specific fine-tuning.

Image-to-character vs Tripo

Text-to-motion vs Hunyuan

Rigging vs Meshy

Video-to-motion vs Cartwheel

What the benchmarks do not show

One checkpoint

Each competitor is specialized for one primitive. Every JST-2 score comes from the same checkpoint, with no task-specific fine-tuning.

Motion is native to the model

JST-2 models space and time jointly, generating motion for each character rather than retargeting it afterward.

From character to motion in seconds

JST-2 produces a motion-ready character in a few seconds without a separate multi-step pipeline. Meshy takes five steps and seven minutes. Tripo takes three steps and fifteen minutes.

Scene generation efficiency

MarbleJST-2 10 min→100 ms $1→under 1¢

Full scene benchmarks are pending the final JST-2 release.

JST-2 specialist

JST-2 demo

Video placeholder

Characters, motion, and scenes from one foundation model

One architecture scales to physical intelligence

Each generation advances the same model across graphics, physics, and reasoning.

JST-1Q4 2022 – Q2 2024 JST-2Q2 2024 – Q2 2026 JST-3Q2 2026 – Q4 2027 JST-XQ4 2027 – Q1 2029 Graphics Proof Solved Solved Solved Physics Basic Simulation-ready Solved Solved Reasoning — Basic Advanced Solved Unlocks A proven architecture and a data flywheel operating at social scale. One foundation model spanning characters, motion, and scenes. A scalable physics-native world simulator that learns rather than follows hand-coded rules. Physical intelligence that reasons, acts, and adapts in simulated worlds at model speed. AI analogue GPT-12018 · architecture proven GPT-22019 · first $1B investment GPT-32020 · scaling breakthrough Beyond LLMsphysical intelligence

JST-1 proved our architecture and data flywheel

50M+ users improved the model through real-world creation and correction, while the product funded the entire loop.

0 → JST-1Architecture proof · complete model stack

TokenizationArchitectureData enginePre-trainingPost-training

JST-1 → JST-2Social-scale data flywheel

A self-sustaining AI flywheel

50M+

users

300M+

videos generated

$2M

ARR

$1.2M

total annual cost

$0

paid acquisition

JST-3: from rule-based to learned physics

The first physics-native world model to learn physical dynamics directly.

JST-2 → JST-3Proven architecture · 10× to 100× model scale

Model-as-a-Service scales to $60M ARR

One foundation model monetized across consumer, developer, and enterprise markets.

Consumer

Products monetize distribution and expand the data flywheel.

Prosumer + developers

Subscriptions and APIs monetize creators, studios, and developers.

Enterprise

Licenses monetize high-volume platform and publisher deployments.

Builders of the first physics-native world model

Four years building the full stack: tokenization, architecture, data engine, pre-training, and post-training.

Hang Chu, CEO

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

Former Principal Researcher, Autodesk AI Lab.
Founding member, NVIDIA Toronto AI Lab.
First author, Modular Codec Avatar, Meta.
PhD ABD, University of Toronto ML Group (led by Geoffrey Hinton); Cornell MS.
2.7K+ citations; h-index 19.

Ming Liang, CTO

Researcher and systems builder for 3D foundation models.

Founding member and staff scientist, Uber ATG.
Core founding member, Waabi.
Founding member, Apple SPG.
Kaggle Data Science Bowl World Champion ($1M, one of Kaggle’s largest-ever prize pools).
Tsinghua PhD; 12.0K+ citations; h-index 31.

Core team

**Jinma**Data LeadTsinghua MS; 10+ years as an ML engineer

QQModel LeadFounding member, SenseTime; 4.0K+ citations

**Yun**Chief ScientistFirst author, CVPR Best Paper Finalist (top 0.2%)

SharpRuntime LeadNVIDIA CUDA team alumnus

**Yao**Infra LeadFormer Senior Engineer, ByteDance

**Jason**Product LeadFormer AI startup founder

**Nan**Growth LeadMaster’s degrees from USC and LSE

KDDesign LeadParsons; formerly at Google

**Eric**Strategy AdvisorFormer CTO, Ctrip; formerly at eBay

**Christine**GTM AdvisorFirst marketer at Twitter

TL;DR

01

Trillion-dollar infrastructure

World simulation will become the trillion-dollar infrastructure layer for physical AI.

02

Physics-native convergence

The field is converging on physics-native foundation models, but the full stack remains unsolved.

03

Generational lead

WarpEngine is the first to train a complete end-to-end physics-native world foundation model.

04

Compounding moat

Our proprietary stack, social-scale data flywheel, and proven scaling capability compound that lead.