Introducing Khora: A Scalable Multiplayer World Model

0:00 / 0:00

Multiple players and agents can act inside one generated world while Khora keeps their views aligned to the same events, locations and consequences.

We built Khora to study a practical question in multiplayer world models: how can a system expand beyond a fixed number of participants without giving each participant a different version of reality?

The public demo allows eight players to join a real-time deathmatch, rendered at up to 32 FPS. The research system is evaluated across multiple layouts and agent populations, with measured runtime experiments from one to 80 agents.

Khora was developed through an ongoing collaboration between Ophilus and RhOS.ai. In our joint work, we explore shared world state, multi-agent consistency and the interaction data needed to train and evaluate systems that act together.

STBoard: one shared world state

Khora uses STBoard (Spatio-Temporal Board), a persistent shared-state architecture. Agent actions are processed by a state-transition model, which updates the shared board before each participant generates a new view from its local conditions and the common world state.

This separates world-state evolution from per-view rendering. Static scene memory remains shared, while dynamic entities keep their own attributes, poses and velocities. A new agent becomes another state entry and view query instead of requiring a new set of pairwise connections.

Khora STBoard architecture showing multi-agent inputs, shared world evolution and per-view rendering — STBoard separates shared world-state evolution from independent per-agent view rendering. Source: Khora technical report (CC BY 4.0).

The architecture is designed around a persistent world rather than a collection of synchronized videos. That distinction is important when one agent changes an object, crosses another agent's path or enters a new area: the consequence must be available to every subsequent view.

Population scaling and linear-time coordination

Direct pairwise communication grows as O(n^2) with the number of agents. STBoard lets each view read the same evolving state, so the dominant coordination path grows approximately linearly at O(n).

Khora population scaling from one to 64 synchronized views with measured latency and FPS — Population scaling with representative 1, 4, 16 and 64-view outputs. Solid measurements and dashed projections are shown separately. Source: Khora technical report (CC BY 4.0).

Selected measured results from the Khora runtime experiment
Metric	Measured result	Scope
Population	1-80 agents	Measured range
Compute latency	107.16-117 ms	Measured runtime
Per-view FPS	37.33-34.27	Measured runtime
Aggregate throughput	2741.3 view-FPS	At 80 agents

These figures describe the paper's measured runtime setup; they are not a promise of unlimited real-world compute. Each additional view still requires rendering resources.

Khora runtime scaling charts for STBoard update, projection, view synthesis, FPS and throughput — Runtime measurements up to 80 agents, with projected trends beyond the measured range. Source: Khora technical report (CC BY 4.0).

Consistency across views and environments

A multiplayer world model must preserve more than visual similarity. It must preserve action consequences, identity, location and causality across multiple viewpoints. Khora tests this through synchronized views, cross-view action consistency, vertical spatial consistency and generalization across maps with different layouts and levels of complexity.

Ancient map used to evaluate Khora generalization — Ancient

Inferno map used to evaluate Khora generalization — Inferno

Mirage map used to evaluate Khora generalization — Mirage

Inference-time population expansion also supports agents joining and leaving a rollout. This makes the system a useful research environment for studying changing multi-agent interaction rather than only fixed multiplayer scenes.

From research system to live demo

The public Khora demo turns these ideas into a playable system. Eight players can join the same generated deathmatch in real time, while the research videos show synchronized two-player, four-player and 16-view runs. The 16-view video demonstrates research-scale rendering; it is not the public demo's player limit.

Khora research run with two players

Khora research run with four players

Khora research run with 16 synchronized views

For us, Khora is a public proof point for a broader data question: training and evaluating interactive AI requires synchronized observations, actions, shared state changes and the outcomes that follow. A video shows what happened; interaction data helps explain why.

Try Khora live demo

Next steps

Our next steps are to extend shared-state generalization across more environments, produce richer multi-agent interaction data, and evaluate long-horizon consistency as populations and behaviors change. We are also expanding the range of gameplay mechanics and shared interactions available in the live system.

The formal methods and full results are available in the Khora technical report. The paper link is coming soon.

Khora is an ongoing collaboration between Ophilus and RhOS.ai. Together, we are continuing to learn from the system, the data it produces and the communities building the next generation of interactive worlds.

Highlight 03 · Scene Transfer