World model API: embed a navigable AI world in your app
Embed navigable AI worlds with a hosted world model API. Compare LingBot, Matrix-Game, DreamX and ABot inputs, movement controls and integration limits.
A world model API lets an application send actions into a generated environment and receive the resulting video. On Windflow, a world runs as a realtime session: provide its supported seed inputs, attach its generated video track, and send movement and look controls while the user explores.
This fits experimental game scenes, interactive demos and visual exploration. Generated video has model-specific continuity and control limits; it should not be treated as a deterministic game engine, collision system or verified physical simulator.
Compare hosted world models by input
| Model | Scene conditioning | Interaction | Starting point |
|---|---|---|---|
| LingBot-World 2.0 | Seed image and text prompt | Persistent movement, look and supported event controls | Prompt-shaped navigable scenes |
| Matrix-Game 2.0 | Seed image; no text prompt | Persistent movement and look | Action-conditioned visual demos |
| DreamX-World-5B | Image and supported prompt conditioning | Movement and look | Comparing a smaller controllable world model |
| ABot-World-0-5B | Seed image; no text prompt | Model-specific movement and look | Third-person exploration |
The catalog publishes each model's inputs, outputs and runtime status. Use those capabilities to decide which controls your interface displays. An empty prompt field on a model that accepts only an image creates a confusing product and cannot add text conditioning.
Build the control loop
- Authenticate the player in your application and issue a short-lived session token from your server.
- Upload an accepted seed image and configure any supported scene conditioning.
- Connect, wait for ready and attach the generated video track.
- Send movement state when a key is pressed or released; send look state from your camera controls.
- Clear movement when the user loses focus, opens a menu or disconnects.
- Release the owner session when exploration ends.
Persistent movement stays active until replaced. Sending idle on key release and window blur prevents a scene from continuing forward after a player stops interacting. Keep local UI feedback immediate, and distinguish a control acknowledgement from the later frame where its effect appears.
Design around the model's limits
Start with a short route through the seed scene. Watch for visual drift, objects changing identity, geometry that fails to stay stable and camera motion that becomes hard to control. Repeat the route with different seeds and controls. A polished single recording cannot establish reliability across users.
Keep game rules, scores, inventory and permissions in your own application. Use the model's video for the visual experience, and validate any product state independently. For research, define what a generated environment can and cannot establish before using observations as evidence.
Plan sessions and cost
Each independent interactive world needs its own owner session and meter. Several authorized viewers can watch one session, but they share that world's state. If every player needs independent navigation, budget independent sessions and peak concurrency.
Windflow is in private beta. GPUs are brought up on demand, so check live readiness and arrange access before a public demo. Current model rates are on pricing. The LingBot hosting guide walks through one model; navigable worlds for games covers product design.
FAQ
Where can I access open-source world models through an API?
Windflow's public catalog includes hosted LingBot-World, Matrix-Game, DreamX and ABot runtimes. Access requires private-beta approval, and live availability depends on GPU capacity.
Can I embed a world model with WASD controls in a browser game?
Yes. Stream the generated video into your browser interface and forward supported movement and look state during the session. Your application should clear movement on key release and focus loss.
Does a generated world provide reliable physics or persistent geometry?
Do not assume it does. Evaluate the model's continuity and action response for your scenario, and keep deterministic game state or safety-critical simulation outside the generated video.