World model release signals
Short release signals, not another model wall.
Read short world model release signals organized by what changed, source confidence, affected models, and next pages to open.
Update signalRelease feed
Scan what changed, source confidence, and where to go next.
The release feed stays short on purpose. Stable descriptions live in Company Map, and choose-between guidance lives in Decision Guides.
High impact
Release signals that changed the site map.
These updates affect the main company map or decision guides and should be read before older timeline notes.
Google expands Project Genie with Street View-grounded worlds
Google connected Project Genie to Street View-grounded scenes and tied the rollout to eligible Google AI Ultra subscribers instead of positioning it as a broad public API launch.
Alibaba launches HappyOyster for real-time world creation
Alibaba positioned HappyOyster as an early-access world-model product for exploration and directing rather than as an open-source model release.
Tencent Hunyuan opens HY-World 2.0 for persistent 3D worlds
Tencent's official HY-World 2.0 repository added HY-Pano 2.0 inference code and weights on May 11, 2026, then added world-generation inference code and WorldStereo 2.0 weights on May 18, 2026.
NVIDIA previews Cosmos 3 as a unifying physical-AI world model
NVIDIA's March 16, 2026 GTC materials linked Cosmos 3 to a broader physical-AI data-factory blueprint, and the public Cosmos GitHub organization remained active through May 27, 2026.
World Labs launches Marble
World Labs launched Marble as a visible product surface for creating, editing, and exploring generated 3D worlds.
Google DeepMind announces Genie 3
Google DeepMind presented Genie 3 as an interactive world-model frontier rather than a fixed clip-generation model.
NVIDIA launches Cosmos
NVIDIA introduced Cosmos as a world foundation model platform for physical AI development.
Watchlist
Release signals that clarify a lane or boundary.
These updates are still source-backed, but they mostly refine a model profile, guide, or availability boundary.
Runway turns GWM-1 into a real-time character product surface
Runway moved GWM-1 from a research framing into a concrete product surface through Characters, with rollout language tied to the Runway API plus web and mobile apps.
Ant Lingguang brings a world-model experience to mobile
Media reports said Ant's Lingguang app added an 'Experience World Model' feature backed by LingBot-World-Fast, but the integration claim still lacks a matching primary Ant product note.
Tencent open-sources HY-Embodied-0.5 and extends it with 0.5-X
Tencent added the HY-Embodied-0.5-X release path on top of HY-Embodied-0.5, widening the robot-facing model surface.
Robbyant releases LingBot-Map for streaming 3D reconstruction
Robbyant added public benchmark and evaluation materials around LingBot-Map, making its streaming 3D reconstruction claims easier to inspect.
Robbyant publishes LingBot-VA for robot-control world modeling
Robbyant published LingBot-VA as a causal video-action world model, creating a clearer robot-control branch next to LingBot-World.
Robbyant releases LingBot-VLA for embodied-AI control
Robbyant published LingBot-VLA as an open vision-language-action release, giving the site a robot-policy signal separate from explorable world simulators.
World Labs announces the World API
World Labs described a programmable World API direction, shifting part of the Marble story from product demo toward developer platform.
Project Sid frames many-agent simulation as an AI civilization experiment
The world-model story expands from generated environments into many-agent behavior, where the world acts as a substrate for social simulation.
Runway introduces GWM-1
Runway introduced GWM-1 as a general world-model direction that reframes its video heritage around worlds, avatars, and robotics-adjacent work.
Oasis demonstrates a realtime playable AI world
Action conditioning became visible in a playable world demo, making the screen feel closer to an interactive environment than a rendered video.
Skybox AI keeps the entry point simple: type a place, open a world
Text-to-world creation became understandable as a 360-degree environment surface rather than only as a flat image or video output.