Model landscape
Video
Models that generate moving images, increasingly with sound. And two systems that generate interactive worlds — which is not the same thing.
Last checked: 10 Oct 2026
What this class is
A video model generates a sequence of frames from text, images or existing footage. The current models add sound in the same pass, accept many references, and edit a clip instead of only generating one.
- What is it optimized for?
- For coherence over time: the same people and objects from frame to frame, plausible motion, and increasingly sound that fits the picture.
- How does it differ from a generalist?
- A generalist describes what happens. A video model has to show it, frame by frame, without contradicting itself.
- What does it output?
- Short clips, often with sound. Interactive world models output a scene that keeps reacting to input.
- Where is it useful?
- Film and advertising drafts, previsualization, explaining a process, editing existing footage. World models: simulated environments for research and training.
Why video is not just many images
A single good frame is not the task. A video model has to hold four things together:
- Coherence over timeA person, an object or a room stays the same from frame to frame.
- MotionThings move the way they would move, not merely from one pose to the next.
- Scene persistenceWhat leaves the frame and comes back is still there, and still the same.
- Spatial and physical consistencySize, depth, light and the behaviour of matter fit together across the clip.
Notable video models
A selection with reasons, not a ranking. Lengths and resolutions are the providers’ own statements.
Video generation
Models that generate and edit clips. None of these providers calls its model a world model.
- Google
Gemini Omni Flash
- Purpose
- Generates videos and edits them in conversation.
- What is notable
- Google's documentation recommends it as the default model for video generation and names Veo 3.1 for particular tasks only. Inputs can be images, audio, video and text.
- Why it is here
- Video as the output of a general multimodal model, not of a separate video model.
- Access
- Through the Gemini API (model ID gemini-omni-1.1-flash).
- Input → Output
- Images, audio, video and text → Video
- Vendor claimGoogle credits the model with an improved intuitive understanding of forces such as gravity and fluid dynamics. We found no independent test of this.
The provider does not call it a world model. Google speaks of "world understanding" but does not call Omni a world model.
Sources: Gemini API docs: Video generation, Gemini API changelog, Google: Introducing Gemini Omni · checked 10 Oct 2026
Model page in the registry: Gemini Omni Flash → - ByteDance
Seedance 2.5
- Purpose
- Generates and edits video and sound together.
- What is notable
- According to ByteDance, 30 seconds of video with sound in one pass, extendable over several rounds.
- Why it is here
- Longer coherent clips with sound from one pass. The provider itself names the physical plausibility of complex motion as an open weakness.
- Access
- Through BytePlus, ByteDance's cloud offering for companies.
- Input → Output
- Text and references → Video with sound
- LimitationByteDance itself names room for improvement in the physical plausibility of complex motion and in scenes with several subjects.
The provider does not call it a world model.
- Kuaishou
Kling 4.0
- Purpose
- Video generation with many references and targeted editing.
- What is notable
- Up to 15 reference assets in one generation: images, videos and subjects, including a specified voice.
- Why it is here
- Control over subjects and scenes through many references rather than text alone.
- Access
- Early access in Kling; on 30 September 2026 the provider named a wider rollout for October.
- Input → Output
- Text plus images, videos and subjects as references → Video
The provider does not call it a world model.
Sources: Kling: Kling 4.0 (30 September 2026) · checked 10 Oct 2026
Model page in the registry: Kling 4.0 → - Black Forest Labs
FLUX 3 Video
- Purpose
- Video with sound from text and images.
- What is notable
- According to the provider, clips of up to 20 seconds with sound from a model that learns image, video and audio jointly.
- Why it is here
- One model for several modalities; the same family also provides the image model.
- Access
- Through the Black Forest Labs API and selected partners. The documentation calls FLUX 3 a preview model.
- Input → Output
- Text and images → Video with sound
- LimitationThe documentation calls FLUX 3 a preview model.
The provider does not call it a world model. Black Forest Labs writes "Real World Models" and "Video Models Must be Reality Models" but does not call the video model itself a world model.
- MiniMax
MiniMax H3
- Purpose
- Video with stereo sound from text, images, video and audio.
- What is notable
- According to the provider, up to 15 seconds at 2K resolution with native stereo sound.
- Why it is here
- A video model with sound whose weights can be downloaded.
- Access
- As a hosted service and, since 3 August 2026, with open weights. The component for 2K output is not part of the open release.
- Input → Output
- Text, images, video and audio → Video with stereo sound
- LimitationThe open weights generate at 768p; the component for 2K is not part of the open release.
The provider does not call it a world model.
Sources: MiniMax: MiniMax H3 (31 July 2026), MiniMax: MiniMax H3 Is Now Open Source (3 August 2026) · checked 10 Oct 2026
Model page in the registry: MiniMax H3 →
Interactive worlds
Systems that generate an environment which responds to input. Google DeepMind calls Genie 3 a world model; Runway uses the word for the predecessor of GWM Worlds 2.
- Google DeepMind
Genie 3
- Purpose
- Generates an environment from a text description that can be explored in real time.
- What is notable
- Not a clip but an explorable scene that responds to input.
- Why it is here
- Interactive world generation is a different task from video generation.
- Access
- Only through Project Genie, for Google AI Ultra subscribers.
- Input → Output
- A text description, then control input → An explorable environment in real time
The provider calls it a world model. Google DeepMind calls Genie 3 a "general-purpose world model".
Sources: Google DeepMind: Genie 3, Google Labs: Project Genie · checked 10 Oct 2026
Model page in the registry: Genie 3 → - Runway
GWM Worlds 2
- Purpose
- Interactive environments generated in real time, with audio.
- What is notable
- According to Runway, continuous 720p video at 24 frames per second with generated audio.
- Why it is here
- A second example of an interactive environment generated in real time, from a different provider.
- Access
- Research preview; Runway itself describes it as such.
- Input → Output
- A description of the scene, then control input → Continuous video with audio
- LimitationRunway writes that real-time generation still trades fidelity for speed.
The provider does not call it a world model. Runway describes the predecessor as a "world model for real-time environment simulation"; the page does not use the word for GWM Worlds 2 itself.
Sources: Runway Research: Introducing GWM Worlds 2 (3 September 2026) · checked 10 Oct 2026
Model page in the registry: GWM Worlds 2 →
No longer available
OpenAI removed the Videos API and the models Sora 2 and Sora 2 Pro from its API on 24 September 2026. The page names no replacement.
Sources: OpenAI API docs: Deprecations · checked 10 Oct 2026
No rankingThe order is the order of this list. agi.jetzt does not grade these models and does not compare them with one another.
Why is this relevant to AGI?
A model that continues a scene plausibly needs some representation of how the world behaves over time: where things are, how they move, what happens next. That is why video models are discussed as a route to richer world representation.
Whether today’s video models have learned such regularities is disputed in research. One study finds their physical understanding severely limited and unrelated to how realistic the video looks; another concludes that scaling alone is not enough to uncover physical laws; a third argues that video models are on a path to becoming general vision models. agi.jetzt takes no side — and uses the term "world model" only where a provider uses it for its own system.
What this does not prove about AGI
A realistic video does not show that a model has learned how the world works. Research finds physical understanding limited even where the video looks right.
Relevance to AGI does not mean demonstrated progress toward AGI. These models are not part of the frontier models, and nothing on this page enters AGI Progress.
Research behind the general statements
- Ho et al.: Video Diffusion Models (2022)Names temporally coherent, high-fidelity video as an important milestone of generative modelling.
- Motamed et al.: Do generative video models understand physical principles? (2025)Finds physical understanding severely limited and unrelated to visual realism.
- Kang et al.: How Far is Video Generation from World Model: A Physical Law Perspective (ICML 2025)Concludes that scaling alone is insufficient for video models to uncover fundamental physical laws.
- Wiedemer et al.: Video models are zero-shot learners and reasoners (2025)The opposing view: video models on a path to unified, generalist vision foundation models.
Sources
All sources read on 10 Oct 2026.
- Gemini API docs: Video generation
- Gemini API changelog
- Google: Introducing Gemini Omni
- ByteDance Seed: Introducing Seedance 2.5 (31 July 2026)
- BytePlus: Dreamina Seedance 2.5 (6 August 2026)
- Kling: Kling 4.0 (30 September 2026)
- Black Forest Labs: FLUX 3 Video, Part 1 (4 August 2026)
- Black Forest Labs: FLUX 3 announcement (23 July 2026)
- Black Forest Labs docs: FLUX 3
- MiniMax: MiniMax H3 (31 July 2026)
- MiniMax: MiniMax H3 Is Now Open Source (3 August 2026)
- Google DeepMind: Genie 3
- Google Labs: Project Genie
- Runway Research: Introducing GWM Worlds 2 (3 September 2026)
- OpenAI API docs: Deprecations