Viggle

Physics-native world foundation model

The first complete end-to-end model stack.

Physical intelligence has a quadrillion-year
experience gap

10⁹ living beings in parallel

×

10⁶ years of evolution

=

10¹⁵ years of embodied experience

Biological intelligence was shaped by it. Physical AI starts with almost none.

Real-robot training

Not a viable path at this scale.

Accurate + efficient world simulator

The only viable path.

The real test of a world model is generality

Generality is a property of the whole stack, not of the output.

Tokenization Architecture Data Engine Pre-training Post-training

JST-2

World Labs / SpAItial AI

Video models

physics-native partial none

Three categories sit outside this table.

JEPA

JEPA reaches for architecture in the abstract while world simulation, the one binding constraint, remains unsolved.

Until it is, the architecture cannot take effect.

Meshy / Tripo

Static shape only. No dynamics or kinematics.

A static asset cannot be driven, so motion capability is absent.

Embodied AI

Slow, expensive, cannot scale.

Robot hardware has not converged. Data collected on one embodiment is written off by the next. The optimal design has to be found by iterating inside a world simulator.

A world simulator has to be accurate and efficient Video models reproduce the phenomenon; physics-native models capture the cause, and cost orders of magnitude less to run.

Model performance against cost

These routes move along the diagonal, not off it: scale up and quality arrives with the bill, scale down and the cost falls with the quality. No engineering budget moves a pixel or reconstruction stack into the upper left, because the limit is the representation, not the implementation.

Positions are qualitative. Axis values pending measured benchmark scores.

What keeps the others on the diagonal

Video generation

Pixels in, pixels out. There is no 3D underneath, so nothing persists between clips and nothing can be precisely controlled. Fidelity only ever arrives with more compute.

Spatial reconstruction

The scene is finished the moment it is made: nothing moves in it, and no one is in it. Move far enough and the geometry runs out.

3D asset generation

Assets, not environments — a scene hundreds of metres across cannot be carried by a mesh at all. Characters and motion are built separately, so the animation never quite fits the character.

Embodied collection

The one route that gathers real physics, at the price of real time and one body at a time. Strong on the task it was collected for; it does not survive the next one.

JST-2 crossed the generality threshold

A world simulator has to cover character, motion and scene. The three cannot be learned apart: motion is how a character interacts with a scene.

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

JST-2 is an all-in-one model across 3 primitives and 6 foundational tasks.

One model, across-the-board SOTA

The same JST-2 checkpoint beats specialist models across all six foundational tasks.

Rigging vs Meshy

Text-to-motion vs Hunyuan

Text-to-scene vs Marble

Image-to-character vs Tripo

Video-to-motion vs Cartwheel

Image-to-scene vs Marble

JST-2 specialist

JST-2 demo

Video placeholder

Scenes + characters + motion, from one model

Needed: final demo video, input prompts, runtime, and output labels.

Our roadmap: from world simulation to experience at model speed

Each generation unlocks the next. Our proven architecture keeps taking on more of the physical world — until it can generate the experience physical intelligence has never had.

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 Game-ready Solved Solved Reasoning Basic Early Growing Solved Unlocks A proven architecture, and a data flywheel turning at world scale. The generality of that same architecture, proven end to end. A scalable, accurate, physics-native world simulator. Learning-based rather than rule-based. An intelligence that acts, fails and adapts inside simulated worlds, at model speed. Value at stake — $10B $100B $1,000B