24 World Models and Adjacent Systems

Last verified: September 4, 2026

Status descriptions are snapshots, not promises of access or performance. Open the linked primary source before choosing a product, API, or research stack.

World creation

Marble

Organization: World Labs
Category: Persistent 3D world creation
First announced: November 12, 2025
Status: Public product with documented Marble 1.1 and API workflows
Summary: Marble turns text, images, video, or spatial layouts into explorable 3D worlds. Its output is a persistent spatial asset rather than a video stream generated only while a session is running.
Boundary: A 3D creation product, not a robotics simulator.
Primary sources: Marble announcement, documentation
Detailed profile: World Models Watch

Skybox AI

Organization: Blockade Labs
Category: 360-degree environment generation
First announced: February 27, 2024
Status: Public product with API and export documentation
Summary: Skybox AI creates panoramic environments for backdrops, VR scenes, lighting, and game-engine workflows. Export options make it useful around a larger 3D pipeline, even though the core artifact is a generated environment panorama.
Boundary: A skybox and environment generator, not a general interactive world model.
Primary sources: Blockade Labs, Skybox export documentation
Detailed profile: World Models Watch

HY-World 2.0

Organization: Tencent Hunyuan
Category: Multimodal 3D world generation and reconstruction
First announced: April 16, 2026
Status: Public code and weights; the official repository now also notes the HY-World 2.1 update
Summary: HY-World combines generation and reconstruction paths for turning text, images, or video into meshes and Gaussian-splat representations. The public repository exposes the technical components rather than only a hosted showcase.
Boundary: Persistent 3D generation and reconstruction, not interactive video simulation.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

World API

Organization: World Labs
Category: Programmable world-generation platform
First announced: January 21, 2026
Status: Public developer interface documented around Marble
Summary: World API exposes World Labs’ world generation as asynchronous developer operations. It returns references to explorable worlds and associated spatial assets for integration into other tools.
Boundary: An API surface for Marble, not a separate foundation model or open-weight release.
Primary sources: World API announcement, API quickstart
Detailed profile: World Models Watch

Echo-2

Organization: SpAItial
Category: Physically grounded 3D world generation
First announced: April 28, 2026
Status: Public application, gallery, and developer API
Summary: Echo-2 converts text, images, or panoramas into spatially persistent 3D scenes that can be explored in a browser. Its public materials emphasize consistent geometry and Gaussian-splat delivery.
Boundary: Current public output is a persistent 3D scene; broader physics and dynamics remain a future direction.
Primary sources: Echo-2 announcement, SpAItial product page
Detailed profile: World Models Watch

SpatialGen

Organization: Manycore Tech and HKUST Spatial AI Lab
Category: Layout-guided indoor 3D scene generation
First announced: August 25, 2025
Status: Public repository, model artifacts, and research implementation
Summary: SpatialGen generates indoor scenes from a 3D semantic layout combined with an image or text-conditioned workflow. It targets consistent room-scale geometry rather than open-ended first-person simulation.
Boundary: Indoor scene generation and reconstruction, not a consumer game or general realtime world.
Primary sources: official repository
Detailed profile: World Models Watch

MoWorld 3D

Organization: KOKONI / MoWorld
Category: Dynamic 3D and 4D scene platform
First announced: July 11, 2026
Status: Officially presented product with limited public self-service detail
Summary: MoWorld is presented as a platform for scene reconstruction, editing, novel views, and dynamic spatial content. Public product information is available, while access, export, and pricing details remain limited.
Boundary: A newly presented platform; public evidence does not yet support treating it as a broadly available self-service creator.
Primary sources: KOKONI MoWorld
Detailed profile: World Models Watch

Interactive worlds

HappyOyster

Organization: Alibaba Token Hub
Category: Real-time multimodal world product
First announced: April 20, 2026
Status: HappyOyster 1.0 gray testing and early-access product surface
Summary: HappyOyster accepts text, voice, and image direction while a user explores or directs a generated environment. Its public positioning combines realtime navigation with scene direction rather than asset export.
Boundary: An early-access creation product, not an open-source model or validated robotics simulator.
Primary sources: HappyOyster, Alibaba Cloud launch article
Detailed profile: World Models Watch

Genie 3

Organization: Google DeepMind
Category: General-purpose interactive world model
First announced: August 5, 2025
Status: Experimental Project Genie access; no public API
Summary: Genie 3 generates environments that a user can navigate in real time and can respond to prompted world events. Project Genie is the experimental interface for creating and exploring these sessions.
Boundary: A research prototype for generated interaction, not a conventional 3D asset pipeline.
Primary sources: Genie 3 model page, research announcement
Detailed profile: World Models Watch

Oasis

Organization: Decart AI and Etched
Category: Action-conditioned interactive video world
First announced: October 31, 2024
Status: Original public demo, code, and weight links; Decart now presents later Oasis generations separately
Summary: The original Oasis project generates a Minecraft-like video world frame by frame in response to keyboard input. It helped make action-conditioned realtime generation tangible through a playable public demo.
Boundary: An interactive video model, not a physics engine or persistent 3D scene editor.
Primary sources: original project page, current Decart Oasis page
Detailed profile: World Models Watch

LingBot-World

Organization: Ant Group / Robbyant
Category: Open interactive world simulator
First announced: January 29, 2026
Status: Public code, paper, and model releases
Summary: LingBot-World is an image-conditioned simulator designed around long-horizon visual consistency and action control. The open repository and checkpoints make the system inspectable beyond a recorded demo.
Boundary: A research stack with substantial local compute requirements, not a polished no-code world product.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

Odyssey-2

Organization: Odyssey
Category: Real-time interactive video world model
First announced: October 27, 2025
Status: Public interactive experience under active product development
Summary: Odyssey-2 streams generated video that changes in response to text during the session. The experience focuses on immediate visual interaction rather than delivering editable scene geometry.
Boundary: A live video world, not a persistent 3D asset.
Primary sources: Odyssey-2 announcement
Detailed profile: World Models Watch

Waypoint 1.5

Organization: Overworld
Category: Local and streamed realtime world model
First announced: April 9, 2026
Status: Browser experience plus open-weight Biome client
Summary: Waypoint generates a first-person visual environment in real time, with options for streamed browser access and local execution on supported hardware. The local path differentiates it from closed browser-only previews.
Boundary: A generated visual environment, not a conventional authored game engine.
Primary sources: Waypoint 1.5 announcement, Biome repository
Detailed profile: World Models Watch

WHAM-RT

Organization: Microsoft Research and Xbox
Category: Real-time world and human action model
First announced: April 4, 2025
Status: Playable Quake II technical demo in Copilot Labs
Summary: WHAM-RT predicts gameplay visuals in response to controller actions quickly enough for an interactive technical demo. It is part of Microsoft’s research into models that jointly represent environments and human actions.
Boundary: A constrained research demonstrator, not a general text-to-game creator.
Primary sources: WHAM-RT, WHAM research overview
Detailed profile: World Models Watch

Matrix-Game 3.0

Organization: Skywork AI
Category: Open streaming interactive world model
First announced: March 27, 2026
Status: Public code and model release
Summary: Matrix-Game 3.0 is a memory-augmented video model conditioned by keyboard and mouse input. Its public repository supports local experimentation with long-form interactive generation.
Boundary: An open research stack with demanding hardware needs, not a supported consumer web game.
Primary sources: official repository, project page
Detailed profile: World Models Watch

Spatial and physical AI

Cosmos

Organization: NVIDIA
Category: Physical-AI world foundation model platform
First announced: January 6, 2025
Status: Active open platform; Cosmos 3 announced in May 2026
Summary: Cosmos provides models and tooling for synthetic data, physical reasoning, world simulation, and action generation. It is aimed at robotics and autonomous systems rather than consumer world creation.
Boundary: A physical-AI development platform, not a single interactive world app.
Primary sources: Cosmos 3 announcement, NVIDIA Cosmos repositories
Detailed profile: World Models Watch

HY-Embodied-0.5

Organization: Tencent Robotics X / HY Vision Team
Category: Embodied foundation model
First announced: April 9, 2026
Status: Public model family, inference code, and weights
Summary: HY-Embodied targets spatial-temporal perception, reasoning, and planning for physical agents. It serves as a reasoning component around robot-control pipelines rather than generating a world for people to explore.
Boundary: An embodied intelligence model, not a visual world creator.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

LingBot-Map

Organization: Ant Group / Robbyant
Category: Streaming 3D foundation model
First announced: April 15, 2026
Status: Public code, paper, checkpoints, and evaluation scripts
Summary: LingBot-Map reconstructs geometry and estimates camera state from streaming video over long sequences. It represents the mapping and spatial-memory side of world modeling rather than generative scene creation.
Boundary: A reconstruction and mapping model, not an explorable-world generator.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

LingBot-VA

Organization: Ant Group / Robbyant
Category: Causal video-action world model
First announced: January 29, 2026
Status: Public code, paper, and model releases
Summary: LingBot-VA models video observations and robot actions together for generalist control. Its purpose is predicting useful action sequences in physical tasks, not rendering environments for a viewer.
Boundary: A robot-control research model, not a consumer simulation product.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

LingBot-VLA

Organization: Ant Group / Robbyant
Category: Vision-language-action foundation model
First announced: January 27, 2026
Status: Public code, paper, and checkpoints
Summary: LingBot-VLA connects visual observations and language instructions to robot actions across multiple hardware configurations. It is included because action models consume and extend spatial representations used in embodied world modeling.
Boundary: A manipulation and instruction-following model, not a standalone environment simulator.
Primary sources: official repository, technical report
Detailed profile: World Models Watch

V-JEPA 2

Organization: Meta
Category: Latent physical reasoning and planning model
First announced: June 11, 2025
Status: Public code, checkpoints, paper, and benchmarks
Summary: V-JEPA 2 learns abstract representations from video for understanding, prediction, and action planning. It predicts in representation space rather than generating a photorealistic world for direct exploration.
Boundary: A latent predictive model for physical reasoning, not a consumer rendering product.
Primary sources: Meta announcement, official repository
Detailed profile: World Models Watch

Research and adjacent systems

EMO

Organization: Alibaba Group, Institute for Intelligent Computing
Category: Audio-driven portrait video research
First announced: February 27, 2024
Status: Public project page, paper, repository, and Model Studio documentation
Summary: EMO animates a portrait from a reference image and vocal audio while preserving identity and expressive motion. It is useful for understanding controllable generated characters, but it does not model an environment around them.
Boundary: Portrait animation, not an explorable or physically predictive world model.
Primary sources: official project page, official repository
Detailed profile: World Models Watch

Project Sid

Organization: Fundamental Research Labs, formerly Altera
Category: Many-agent civilization simulation
First announced: November 1, 2024
Status: Public research report and repository
Summary: Project Sid studies large groups of autonomous agents developing roles, rules, and culture inside Minecraft. The environment is authored; the research contribution is the behavior of the agents within it.
Boundary: An agent-society simulation, not a generative visual world model.
Primary sources: Project Sid report, official repository
Detailed profile: World Models Watch

GWM-1

Organization: Runway
Category: General world model research program
First announced: December 11, 2025
Status: Official research program with product-facing character work
Summary: GWM-1 frames Runway’s work across interactive worlds, realtime characters, and robotics-oriented simulation. Public materials describe a family and direction rather than one downloadable general-purpose model.
Boundary: A research and product program; access and capabilities differ across its component surfaces.
Primary sources: GWM-1 research page, Runway Characters article
Detailed profile: World Models Watch

For comparisons, practical generation workflows, and ongoing release tracking, visit the World Models Watch model directory.