← Guides
Model hosting guideUpdated 3 min readBy Windflow Labs

How to host LingBot-World as a realtime API

Run LingBot-World through a hosted API or evaluate self-hosting. Learn seed inputs, persistent movement, session costs and checks for an interactive app.

Windflow hosts LingBot-World 2.0 as a realtime session API. A developer supplies a seed image and text prompt, receives a navigable video stream, and sends persistent movement and look state while it runs. Access is through private beta, with GPU capacity brought up on demand.

The upstream model is from Robbyant. The official LingBot-World repository is the source for research code and its license information. The Windflow model page is the source for Windflow's serving inputs, output specification, pricing and runtime status.

Hosted API or your own serving layer

A hosted session is useful when you want to evaluate interaction inside a product and meter ready time without operating the worker. Self-hosting gives control over weights, infrastructure and runtime configuration, and requires operating the streaming and session layers yourself.

For self-hosting, pin the code revision and weights, verify the license for your intended use, establish a repeatable inference baseline, and add browser transport, session lifecycle, authentication and observability. A working research demo does not establish a production API. Compare the operational trade-offs in managed hosting versus rented GPUs.

Prepare a minimal interactive experience

Start with one scene and one player. Choose a seed image with a clear perspective and route, write a short scene prompt, and build the player around the generated video track. Keep the initial control surface limited to movement, look and an explicit end-session action.

Use your backend for account authentication and session-token issuance. Keep your Windflow account key off the browser. Once the session is ready, apply the supported scene inputs and follow the model's start flow from the authenticated API documentation.

Handle persistent movement correctly

LingBot's movement control expresses state, rather than a single animation tick. Set a direction when the user begins moving, replace it when direction changes, and send idle when movement ends.

  • Clear movement on key release and when the browser loses focus.
  • Give menus and text fields priority over navigation keys.
  • Use supported look controls rather than inventing an unsupported camera command.
  • Display control failures, including invalid input or a session that has ended.
  • Observe the delay from command to visible response on your actual network.

The acknowledgement says the command arrived; it does not prove the user has seen its effect. The latency guide separates those measurements.

Decide whether the scene is good enough

Run the same route several times and record where continuity breaks. Include reversing direction, turning to revisit an object and stopping after sustained movement. Inspect stability across a longer session rather than judging only the opening seconds.

If you need image-only conditioning, compare Matrix-Game 2.0. If you want to compare another prompt-conditioned model, evaluate DreamX-World-5B using the same scene and route.

Readiness and billing

Windflow's catalog status and current readiness answer different questions. A production runtime can have no free GPU at the moment you connect. Design queued, offline and reconnecting states, and arrange a live demo before a time-sensitive event.

Use the rate on the LingBot model page or pricing. Queueing and startup are free; ready time is billable even if the player stops moving. Release the owner session when done.

FAQ

Can I run LingBot-World without managing a GPU?

Yes, through Windflow's hosted LingBot-World 2.0 session API after private-beta approval. Live capacity is separate from the catalog's implementation status.

Can I change the environment with a text prompt?

LingBot supports prompt conditioning and model-specific event controls. Check the published model capabilities and the authenticated command reference for the controls and timing available to your session.

Does stopping movement stop billing?

No. Stopping movement leaves the session ready and the GPU reserved. Release the owner session when the experience ends.

Run realtime models inside your product.