{ "schemaVersion": "1.0.0", "updatedAt": "2026-07-18", "site": "https://worldmodelatlas.com/", "pageUrl": "https://worldmodelatlas.com/pulse/", "methodUrl": "https://worldmodelatlas.com/method/", "coverage": { "label": "Selected public activity from tracked world-model teams", "lastSourceReview": "2026-07-10", "isExhaustive": false, "note": "This first edition is a monitoring baseline built from official releases and primary research already reviewed by World Model Atlas. It is not a complete news feed or an independent capability ranking.", "trackedTeams": [ "Google DeepMind", "NVIDIA", "Meta AI", "OpenAI", "World Labs", "Wayve", "Waabi", "Runway" ] }, "entries": [ { "id": "nvidia-cosmos-3-2026", "teamId": "nvidia", "team": "NVIDIA", "project": "Cosmos 3", "publishedAt": "2026", "datePrecision": "year", "retrievedAt": "2026-07-10", "type": "model_release", "title": "Cosmos 3: Omnimodal World Models for Physical AI", "summary": "NVIDIA Research presents Cosmos 3 as an omnimodal world-model system spanning language, video, action and audio signals for physical-AI settings.", "whyItMatters": "It expands the Cosmos route from a video-generation component toward a broader modeling and tooling stack for robots and autonomous systems.", "sourceRole": "primary_research", "evidenceNote": "Developer-authored technical report; it does not independently validate an end-to-end robot or driving system.", "sourceId": "nvidia-cosmos-3", "sourceTitle": "Cosmos 3: Omnimodal World Models for Physical AI", "sourceUrl": "https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf", "media": [], "tags": ["physical AI", "omnimodal", "robotics", "autonomous driving"] }, { "id": "deepmind-genie-3-2025", "teamId": "google-deepmind", "team": "Google DeepMind", "project": "Genie 3", "publishedAt": "2025", "datePrecision": "year", "retrievedAt": "2026-07-09", "type": "model_release", "title": "Genie 3: A new frontier for world models", "summary": "Google DeepMind describes Genie 3 as a general-purpose world model that generates dynamic environments from text and supports real-time navigation.", "whyItMatters": "The release makes interactive, action-responsive generated worlds a more visible research direction beyond fixed video output.", "sourceRole": "official_release", "evidenceNote": "Official capability claims; public access and independent long-horizon evaluation remain limited.", "sourceId": "deepmind-genie-3", "sourceTitle": "Genie 3: A new frontier for world models", "sourceUrl": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/", "media": [ { "type": "demo", "label": "Official interactive-world demonstrations", "url": "https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/" } ], "tags": ["interactive worlds", "text-to-world", "agent training"] }, { "id": "meta-vjepa-2-2025", "teamId": "meta-ai", "team": "Meta AI", "project": "V-JEPA 2", "publishedAt": "2025", "datePrecision": "year", "retrievedAt": "2026-07-09", "type": "research", "title": "Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning", "summary": "Meta presents V-JEPA 2 as a predictive representation model trained from video, with evaluations for physical reasoning and a separate action-conditioned robot-control route.", "whyItMatters": "It represents a world-model strategy centered on latent prediction and planning rather than direct photorealistic generation.", "sourceRole": "official_release", "evidenceNote": "Official research release with developer-run benchmarks; task transfer and protocol limits remain important.", "sourceId": "meta-vjepa-2", "sourceTitle": "Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning", "sourceUrl": "https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/", "media": [ { "type": "demo", "label": "Official research demonstrations", "url": "https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/" } ], "tags": ["JEPA", "physical reasoning", "robot control", "representation learning"] }, { "id": "world-labs-marble-2025", "teamId": "world-labs", "team": "World Labs", "project": "Marble", "publishedAt": "2025", "datePrecision": "year", "retrievedAt": "2026-07-09", "type": "product", "title": "Marble: A Multimodal World Model", "summary": "World Labs presents Marble as a multimodal system for creating, editing, expanding and exporting 3D worlds.", "whyItMatters": "It moves the product question from generating a clip toward creating spatial environments that can be explored, modified and reused in other workflows.", "sourceRole": "official_release", "evidenceNote": "Official product positioning and demonstrations; physical fidelity and downstream usefulness require separate evaluation.", "sourceId": "world-labs-marble", "sourceTitle": "Marble: A Multimodal World Model", "sourceUrl": "https://www.worldlabs.ai/blog/marble-world-model", "media": [ { "type": "demo", "label": "Official product demonstrations", "url": "https://www.worldlabs.ai/blog/marble-world-model" } ], "tags": ["spatial intelligence", "3D worlds", "world generation"] }, { "id": "wayve-gaia-2-2025", "teamId": "wayve", "team": "Wayve", "project": "GAIA-2", "publishedAt": "2025", "datePrecision": "year", "retrievedAt": "2026-07-10", "type": "research", "title": "GAIA-2 technical report", "summary": "Wayve reports a controllable, multi-view generative world model conditioned on driving actions, road semantics and environmental factors.", "whyItMatters": "It shows how a world model can be specialized for scenario generation and controlled rollouts in autonomous-driving research.", "sourceRole": "primary_research", "evidenceNote": "Developer-authored technical report; generated scenarios are not equivalent to independently calibrated policy-in-the-loop safety validation.", "sourceId": "wayve-gaia-2", "sourceTitle": "GAIA-2 technical report", "sourceUrl": "https://arxiv.org/abs/2503.20523", "media": [], "tags": ["autonomous driving", "controllable generation", "multi-view"] }, { "id": "openai-sora-world-simulators-2024", "teamId": "openai", "team": "OpenAI", "project": "Sora", "publishedAt": "2024", "datePrecision": "year", "retrievedAt": "2026-07-09", "type": "research", "title": "Video Generation Models as World Simulators", "summary": "OpenAI's technical note frames large-scale video generation as a possible path toward general-purpose simulators of the physical world.", "whyItMatters": "It helped make the boundary between video generation and world modeling a central public research question.", "sourceRole": "official_release", "evidenceNote": "Official technical framing; it does not establish persistent action-conditioned interaction or planning utility.", "sourceId": "openai-sora-world-simulators", "sourceTitle": "Video Generation Models as World Simulators", "sourceUrl": "https://openai.com/index/video-generation-models-as-world-simulators/", "media": [ { "type": "demo", "label": "Official generated-video examples", "url": "https://openai.com/index/video-generation-models-as-world-simulators/" } ], "tags": ["video generation", "world simulators", "physical consistency"] } ] }