The realtime world cloud · Private beta

Build worldsthat answer back.

Windflow turns interactive video and world models into continuous, controllable sessions. Connect once, receive live frames, and send the controls each model supports while it runs.

01Why sessions

A request ends. A world session continues. Connect once, receive live frames, and steer what happens next while the model is still running.

Persistent sessions

Model state stays alive across commands, frames, and recoverable reconnects, so an interactive world can continue without rebuilding context.

Bidirectional control

Prompts, images, camera motion, and actions travel over the same live control channel, reshaping the active stream without restarting it.

GPU runtime

Long-lived GPU workers keep model state warm while WebRTC returns continuous frames through a low-latency session path.

Stateful metering

Usage begins when the runtime is ready and is recorded against actual session seconds, keeping every live world tied to a clear lifecycle.

02Model systems

Different models,
one session boundary.

4 of the 10 production model runtimes are shown here. Capacity is reported separately from model lifecycle status.

Full model catalog

WF / verified signals

Current evidence

Private-beta scope. Measurements are internal and not an SLA.

Model runtimes

10 production

Current

Developer SDKs

JavaScript · Python

Current

Reference applications

Generation · editing · worlds · multi-shot

Current

Faster Helios readiness

Compile cache · about 12 to 2 min

Windflow measured

Active system

Frames, state, and control in one loop

WF / 03
03Current architecture

One session.
Every signal.

One session boundary connects the developer contract, lifecycle control, GPU execution, and live delivery across the whole runtime.

DeveloperInteractive Session Protocol · JavaScript + Python
ControlAccess · lifecycle · session-second usage ledger
ComputeModel adapters · GPU runtime · model-specific state
MediaWebRTC video · reliable data commands
04Session mechanics

The first frame
is the easy part.

Interactive worlds break the scheduler, the cost model, and the request lifecycle of an inference cloud. Four mechanisms replace them.

WF / 01 · DeadlineInstrumented
set_prompt

Generation · 33-frame blocks

in-flight retirednew intent

Delivery

Frames owe deadlines. Controls land at block edges.

Every command lands at the next generation-block edge: a two-level cancel retires the in-flight block, and delivery never gaps.

51 ms budget · 33-frame blocks

Windflow measured · private beta · not an SLA

05One Session API

Commands in.
Live frames out.

Create a session, subscribe to video and state, then send supported controls without restarting it. Model-specific inputs and limits remain explicit.

WF / Session API
Live transport
session.tsTypeScript
01
02
03
04
05
06
07
08
09
10
11
12
13
14
15
import { Windflow } from "@windflow/sdk";const session = new Windflow.Session({  model: "helios",});session.on("trackReceived", ({ name, track }) => {  if (name === "main_video") video.srcObject =    new MediaStream([track]);});await session.connect(jwt);await session.sendCommand("start", {  prompt: "a city unfolding at dawn",});
Realtime SDK · Private betaConnected transport
Session shapeStateful
01

Create

Allocate a model runtime

02

Connect

Attach realtime transport

03

Control

Send supported commands while live

04

Observe

Read frames, state, and usage

06Get started

Put a world
in motion.

Building an interactive video product or bringing a new world model online? Join the private beta and test it against a real session workflow.