Persistent sessions
Model state stays alive across commands and reconnects, so a world keeps going instead of starting over.
Realtime video editing · Continuous video · Interactive worlds · Audio + video
Realtime video.Interactive worlds.
Run realtime video and interactive world models inside your product: one API, streamed live, billed by the second. 10 models for video and interactive worlds.
01Why sessions
A request ends. A session keeps running. AI video is moving from finished clips to live streams that respond to every input, and it needs a platform built for that.
Live control
Prompts, images, camera, and movement change the stream while it plays, on the same connection.
Continuous streaming
Long-lived sessions keep each model ready and stream its frames over WebRTC.
Per-second billing
Billing starts when the session is ready, never while it queues or boots, and stops when it is released.
02Model catalog
10 models.
One API.
Video editing, continuous video, interactive worlds, and audio with video, all behind the same API. Every model page lists inputs, controls, price, and limits.
Full model catalog
WF / 01
Recorded sample
Runtime statusProduction
$0.0017Per ready second
Steerable video
Helios
Video that keeps going from a prompt or an image. Change the prompt and the scene follows.
- In
- Text or an image
- Control
- New prompts mid-stream
- Out
- One continuous video stream
Production· $0.0017 / s
33-frame generation chunks
- In
- Text or an image
- Control
- New prompts mid-stream
- Out
- One continuous video stream
View model
WF / 02
Recorded sample
Runtime statusProduction
$0.0021Per ready second
Multi-shot direction
LongLive 2.0
Direct a scene shot by shot: ease into the next shot, or cut to a new one.
- In
- Prompt + optional image
- Control
- Next shot · soft change or cut
- Out
- A directed sequence
Production· $0.0021 / s
1280 × 704 · multi-shot
- In
- Prompt + optional image
- Control
- Next shot · soft change or cut
- Out
- A directed sequence
View model
WF / 03
Recorded sample
Runtime statusProduction
$0.0019Per ready second
Live video effects
SANA-Streaming
Restyle a webcam or a clip as it plays. The motion stays; the look changes.
- In
- Webcam or uploaded clip
- Control
- A text instruction for the style
- Out
- The same motion, restyled live
Production· $0.0019 / s
1280 × 704 · native edit
- In
- Webcam or uploaded clip
- Control
- A text instruction for the style
- Out
- The same motion, restyled live
View model
WF / 04
Recorded sample
Runtime statusProduction
$0.0024Per ready second
Explorable worlds
LingBot-World 2.0
Turn one image into a world, then walk through it: move and look around as it renders.
- In
- One seed image
- Control
- WASD movement + look
- Out
- A world you can walk through
Production· $0.0024 / s
1408 × 768 delivery
- In
- One seed image
- Control
- WASD movement + look
- Out
- A world you can walk through
View model
WF / verified signals
Measured
Measured by Windflow. Methods and raw runs are in the linked report.
Video with sound, faster than realtime
MiniMax H3: a 14.4 s window in 9.30 s · 960×544
Windflow measuredJoyAI live edit, camera frame to edited frame
p90 578 ms · 1248×720 at 24 fps
Windflow measured · medianDeveloper SDKs
JavaScript · Python
Private beta
Full stack
Model runtime to product, one platform
03The stack
Model to product,
one stack.
Windflow runs every layer between a realtime model and your product, so your team builds the experience, not the infrastructure.
Model runtimeContinuous generation · model state kept across commands
Session controlAccess keys · lifecycle · per-second billing
Live deliveryWebRTC video · commands on a reliable channel
API + SDKOne session API · JavaScript and Python SDKs
AppsPlayground · sample apps · your product
04Why rebuild
The first frame
is the easy part.
Inference clouds were built for request and response. Realtime models behave like a video call: one continuous session, state kept, frames never stopping. We rebuilt that layer.
WF / 01 · DeadlineIllustration
sendCommand · set_prompt
Generation · 33-frame blocks
in-flight retirednew intent
Delivery
no gapFrames never wait for a new instruction.
On block-based models such as Helios, a new prompt lands at the next block. The block in progress is dropped, and video keeps playing while the next one renders.
Helios · 33-frame blocks
05One session API
Commands in.
Live frames out.
Connect once, play the stream, send commands. The same session code works for every model; only the commands change.
Same session · any model
session.tshelios
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") return; if ("attach" in track) track.attach(video); else video.srcObject = new MediaStream([track]);});await session.connect(jwt);await session.sendCommand("set_prompt", { prompt: "a city at dawn" });await session.sendCommand("start", {});// Steer it while it runs.await session.sendCommand("set_prompt", { prompt: "the same city at night" });
Realtime SDK · Private betaExample code
Session shapeFour steps
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
Build with realtime video and world models
Windflow API: common questions
- What is Windflow?
- Windflow runs realtime interactive AI video models as on-demand sessions. Teams plug them into their products and pay by the session second.
- Which realtime video and world models can I run through the API?
- The public catalog includes Helios, LongLive-2.0, SANA-Streaming, JoyAI-Video-Edit, LingBot-World 2.0, Matrix-Game 2.0, DreamX-World-5B, ABot-World-0-5B, MiniMax FastH3, LTX-2.5. Choose by supported inputs, controls and outputs. A published runtime status describes implementation maturity; live capacity is reported separately.
- How is a realtime video API different from a video generation job?
- A realtime session keeps model state alive, delivers a video track and accepts supported controls while it runs. A video generation job takes an input and returns a completed result. Windflow provides sessions for live editing, continuous generation and navigable worlds.
- How much does the Windflow API cost?
- Public model rates range from $0.0017 to $0.0069 per ready session-second. Queueing and worker startup are free. A ready session remains billable when its video is paused or unwatched; release it to stop the meter.
- Can I embed an interactive world model in my game or app?
- Yes. Use a model that accepts movement and look controls, stream its generated video into your interface, and forward the user's controls during the session. Windflow offers JavaScript, React and Python integrations. Generated worlds have model-specific continuity limits and do not provide a deterministic physics engine.
- How can I try Windflow or bring my own model?
- Windflow is in private beta. Request access or a live demo; approved accounts receive $5 and 4 concurrent sessions. Model teams can submit licensed weights and a reproducible inference path through Bring a model.
06Get started
Bring your product.
Or your model.
Building with realtime models? Approved accounts start with $5 and pay only for ready seconds. Training one? We make it stream live and put it in front of developers on the same platform.
No invite yet? Request one by email