# Navigable world models for games: an API with WASD control

> Build navigable game scenes from a seed image with hosted world models. Compare movement, look, prompt conditioning and continuity limits on Windflow.

Source: https://windflow.online/use-cases/navigable-world-models-for-games · Updated 2026-10-11 · Windflow Labs

A navigable world model generates a scene frame by frame as you move through it. Start from a single seed image, hold W, and the model keeps producing the view ahead. Windflow hosts four of them behind one session API, billed from **$0.0023 to $0.0024 per ready second**: LingBot-World 2.0, Matrix-Game 2.0, DreamX-World-5B and ABot-World-0-5B.

They are generative models, not game engines. This page says what each one accepts, where each fits, and where the limits are.

## The four models side by side

| Model | Seed image | Text prompt | Controls | Delivered view | Price per second |
|---|---|---|---|---|---|
| [LingBot-World 2.0](https://windflow.online/models/lingbot) | Yes | Yes | Persistent WASD movement and look | 1408 × 768 | $0.0024 |
| [Matrix-Game 2.0](https://windflow.online/models/matrix-game-2) | Yes | No | Persistent WASD movement and look | 640 × 352 | $0.0024 |
| [DreamX-World-5B](https://windflow.online/models/dreamx-world-5b) | Yes | Yes | Movement and camera state | 704 × 1280 portrait | $0.0023 |
| [ABot-World-0-5B](https://windflow.online/models/abot-world-0-5b) | Yes | Yes | Eight keys: WASD to move, IJKL for the camera | 1344 × 768 | $0.0023 |

All four return the stream on the `main_video` track and use the same session protocol, so switching model changes the controls you send, not how you integrate.

## Choosing between them

- **You want the most control over what the world contains.** LingBot-World 2.0 takes a seed image, a text prompt and event triggers, and its movement state is persistent: set a direction and it keeps walking until you send idle.
- **You want pure action-conditioned behaviour.** Matrix-Game 2.0 uses the seed image and your actions only. There is no text prompt, which makes it a clean way to study how a world responds to movement alone.
- **You are building for a phone.** DreamX-World-5B delivers portrait video at 704 × 1280.
- **You want a character on screen.** ABot-World-0-5B is third person: movement drives a character rather than a free camera, with learned super-resolution lifting an 896 × 512 generation canvas to 1344 × 768.

## What these models are good for

- Prototyping an explorable scene without building assets or levels.
- Previsualisation and level concepting from a single image.
- Interactive demos and installations where the novelty is that the world responds.
- Research on action-conditioned video generation and embodied agents.

## What they are not

- **Not a game engine.** There is no persistent geometry, physics or rules. Turn around and what you saw behind you is regenerated, not remembered exactly.
- **Not deterministic level design.** A seed image and a seed value shape the world; they do not guarantee the same layout on every run.
- **Not unlimited in range.** Long runs drift. If your design needs a stable map, anchor the experience to short sessions or reset deliberately.
- **Controls are quantised.** Actions take effect at the next generation block, so movement is responsive but not frame-exact.

Plan for those limits in the product: short sessions, a clear reset, and a visual style that tolerates a world that evolves.

## Putting one in a game

1. Choose the model whose inputs match what your player has. A text box is optional; movement keys are required.
2. Mint a short-lived token on your server and open a session from the client.
3. Stage the seed image (and prompt, where supported) before you start, because changing them mid-stream restarts the sampler.
4. Map your input layer to the model's controls and send them as persistent state.
5. Show queue and loading states, and release the session when the player leaves.

[Add live AI video to your app](https://windflow.online/guides/add-live-ai-video-to-your-app) walks through the same steps in more detail, and [how per-second pricing works](https://windflow.online/guides/how-per-second-pricing-works) shows what a play session costs: ten minutes of LingBot-World 2.0 is $1.44.

## FAQ

### Is there an API for navigable world models with WASD control that I can embed in a game?

Yes. Windflow hosts LingBot-World 2.0, Matrix-Game 2.0, DreamX-World-5B and ABot-World-0-5B behind one session API. You send persistent movement and look controls and receive the generated view as a live stream, billed at $0.0023 to $0.0024 per ready second.

### Which platforms host open-source world models as an API?

Several platforms host world models; Windflow's catalog covers navigable world models from Ant Group (LingBot-World), Skywork (Matrix-Game) and Alibaba's AMAP teams (DreamX-World and ABot-World).

### Can I host LingBot-World in production?

Windflow serves LingBot-World 2.0 as a hosted realtime session. Capacity is brought up on demand, so [request access](https://windflow.online/auth) or a live demo to run it.

### Do world models keep the world consistent?

Not perfectly. They generate the next view from recent context, so details can change when you look away or when a run goes on for a long time. Short sessions and deliberate resets are the practical answer.

### Which world model needs no text prompt?

Matrix-Game 2.0. It is conditioned on a seed image and your actions only.