Models

Agora-1

Odyssey

Multi-agent world model for multiple human or AI participants in one real-time simulation.

Alaya-EVOKE

Alaya Lab

Alaya Lab’s successor to AlayaWorld — an interactive world model that externalizes persistent world state (a camera-indexed world state bank) to sustain long, coherent sessions, and supports re-prompting mid-generation without restarting. Open-weight (Apache 2.0), state-of-the-art on the WBench interactive world-model benchmark as of August 2026.

AlayaWorld

Alaya Lab

Interactive autoregressive world model with real-time camera control, prompt switching, and long-horizon memory consistency. Open-source, sustains interactive play past the one-minute mark.

Atlas

World Labs

World Labs' omni world model — a multimodal autoregressive diffusion transformer pretrained natively on text, images, video, and 3D. Generates up to a minute of camera-controlled 1440p video, reconstructs explicit 3D geometry (point clouds and Gaussian splats), and simulates robot sensor views for Real-to-Sim workflows. In early access with select partners; World Labs says it will power future versions of Marble.

Cosmos 3

NVIDIA

Omnimodal world-model family for physical AI spanning language, image, video, audio and action.

Genie 3

Google DeepMind

General-purpose world model generating photorealistic environments from text with real-time exploration.

1 world on QuantWorlds

HY-World 1.5 (WorldPlay)

Tencent Hunyuan

Tencent Hunyuan's streaming video diffusion model for real-time, interactive world modeling with long-term geometric consistency, released as WorldPlay. Given a single image or text prompt, it generates a next-chunk (16-frame) video prediction conditioned on live user keyboard/mouse actions, dynamically reconstituting context memory from past frames to keep the world geometrically consistent over long sessions. Runs at up to 24fps in first- or third-person, across real-world and stylized scenes. Open-sourced (code, weights, and full training pipeline). Predecessor to, and released separately from, the later HY-World 2.0.

1 world on QuantWorlds

HY-World 2.0

Tencent Hunyuan

Multi-modal world model framework for world generation and world reconstruction — accepts text, single-view images, multi-view images, and video to produce editable, persistent 3D world representations (meshes / 3D Gaussian Splatting) compatible with game engines (Blender, Unity, Unreal).

Khora

Ophilus

A scalable multiplayer world model for shared AI worlds, built by Ophilus in collaboration with RhOS.ai. Supports arbitrary player populations with linear-time complexity while keeping a coherent shared world across players during fast movement, elevation changes, and combat. Whether the demo's map geometry is generated live per-session, or a fixed environment streamed in real time, has not been independently confirmed — the live multiplayer interactivity itself has been.

1 world on QuantWorlds

LTX-2.5

Lightricks

Lightricks' 22-billion-parameter open-weights audio-video generation model, released August 11, 2026. Turns text, image, and video inputs into synchronized, high-fidelity video and audio in a single pass, including native multishot generation (connected scenes with consistent character/style across cuts). Positioned by Lightricks and press coverage as a "world model" for video, robotics and simulation, distinct from this catalogue's live-interactive world models (Marble, MIRA, PAN, WorldPlay) — see research notes for the classification nuance.

Lucy 2.5

Decart

Real-time world editing/transformation model for live immersive video experiences.

Marble 1.0 Draft

World Labs

Fast Marble variant for rapid exploration.

Marble 1.1

World Labs

World Labs world-generation model with improved quality at fixed generation cost.

1 world on QuantWorlds

Marble 1.1 Plus

World Labs

Advanced Marble model from World Labs for larger persistent 3D worlds.

Marionette

Alaya Lab

Marionette predicts an explicit, interpretable 3D world state, renders its geometry with a graphics operator that has no learnable parameters, and asks a video-diffusion model for one thing only: appearance. A world model for interactive games with articulated characters, trained on recordings from a commercial action game (drawn from the WildWorld corpus).

Matrix-Game 3.0

Skywork AI

Real-time and streaming interactive world model with long-horizon memory, for controllable game world generation.

MIRA

General Intuition

Multiplayer Interactive World Models with Representation Autoencoders — a 5-billion-parameter diffusion transformer (with a 600M-parameter video representation codec) that simulates 2v2 Rocket League matches in real time at 20 FPS, taking action streams from up to 4 agents at once. Trained on ~10,000 hours of bot-generated 2v2 matches, without an explicit physics engine or 3D representation.

1 world on QuantWorlds

Oasis 3

Decart

Interactive world model for physical AI with controllable multi-view simulation in real time.

Odyssey-2

Odyssey

General-purpose real-time world model generating interactive simulations from text or image prompts.

An open reproduction of DeepMind's Dreamer4 world-model pipeline, built by researchers Diego Marti Monso, Francesco Sacco and Edward Hu with infrastructure support from Reactor. A 1.6B-parameter dynamics model (30 block-causal transformer layers, JAX/Flax NNX) trained on Minecraft/VPT-style gameplay data, generating real-time interactive video frames from live keyboard and mouse input.

1 world on QuantWorlds

PAN

MBZUAI

General, interactable, long-horizon world simulation model — can be manipulated at intermediate steps and maintains consistency over long time horizons, evaluated as competitive with leading commercial world models.

1 world on QuantWorlds

Starchild-1

Odyssey

Odyssey world model exploring richer multimodal interaction beyond visual observation alone.

Waypoint-1.5

Overworld

Overworld's real-time video world model for consumer GPUs: a 1.2B-parameter autoregressive diffusion transformer generating interactive 720p/60fps video from keyboard and mouse input, with a 512-frame context window. Released open-weight under Apache 2.0 on Hugging Face in 1B and 1B-360P tiers. The hosted overworld.stream platform requires an account to play; the model weights themselves do not.