A catalog entry
A published lifecycle state, supported controls, serving profile, and known limits on the Models page.
For model teams
Bring your model to developers.
Windflow ports real-time models to a production runtime on NVIDIA and AMD GPUs, optimizes inference, wraps them in one Session API, and puts them in front of developers with numbers we measured ourselves. Non-exclusive. You keep your rights.
01 / What you get
A runtime, users, and numbers we can both defend.
Real sessions
The same API, Playground, and hosted MCP path developers already use across the catalog.
Numbers we can both defend
Windflow-measured FPS, first-frame time, continuity, and control latency, reviewed with you before anything goes public.
How evaluation works · typical production launch
2–4
weeks from a complete handoff
0030 minutesIntro call
011–2 working daysCheck the handoff
023–7 working daysReproduce + package
033–7 working daysIntegrate the session
042–5 working daysJoint acceptance
Custom kernels, multi-GPU topology, license review, or an unstable upstream path pause or extend the clock. We say so at the handoff check, not after.
02 / Three ways to work together
Pick the path that matches how your model ships today.
Every path ends in the same session contract. Open weights are the standard path today; closed models come in through an evaluated endpoint or a scoped project.
Path / 01Available now
Licensed weights, hosted by Windflow
- You provide
- A checkpoint at an exact revision, hosting rights, a runnable path, and a measured baseline.
- Windflow does
- Packaging, optimization, serving on NVIDIA and AMD, a catalog entry, and developer access.
Path / 02By arrangement
Vendor API integration
- You provide
- An inference endpoint, rate limits, a latency profile, and commercial terms.
- Windflow does
- An evaluation of latency and controls, then your endpoint behind the same session contract.
Path / 03By arrangement
Joint delivery
- You provide
- A scoped goal: a new GPU vendor, a custom serving profile, or a co-launched product.
- Windflow does
- A project plan with milestones and joint acceptance.
Model team
Complete handoff
Weights or an endpoint, rights, launch path, and baseline. In every path, the runtime work is ours.
Runtime image
Pin source and dependencies, verify weight restore, make startup repeatable, and expose health.
Session contract
Map readiness, lifecycle, reconnects, named tracks, messages, limits, and failures.
Controls + media
Wire only the prompts, images, video, movement, camera, or actions the model really supports.
Recording + usage
Attach recording lifecycle, observability, and ready-time metering to the public session path.
Result
Public session
Measured, documented, recordable, and metered.
Protocol proof / JavaScript SDK
Commands in. A named media track out.
This is the production session surface: a typed command, a reliable control channel, and the generated video on a recvonly track.
- Command
- set_prompt · prompt: string
- Track
- main_video · recvonly · video
import { Windflow } from "@windflow/sdk";
const session = new Windflow.Session({ model: "helios" });
session.on("trackReceived", ({ name, track }) => {
if (name !== "main_video") return;
// The LiveKit path hands over a RemoteTrack; the WebSocket path a MediaStreamTrack.
if ("attach" in track) track.attach(video);
else video.srcObject = new MediaStream([track]);
});
await session.connect(jwt);
await session.sendCommand("set_prompt", {
prompt: "Misty avenue at dawn",
});
await session.sendCommand("start", {});03 / The licensed-weights path
Four things before GPU work begins.
These apply when Windflow hosts your weights. If one is missing, we say so at the handoff check before scheduling compute. The vendor API path needs an endpoint, rate limits, a latency profile, and commercial terms instead.
Weights
A complete checkpoint at an exact revision.
We need every required model asset and a stable way to fetch it during packaging and acceptance.
Evidence · Checkpoint or registry path · source revision · asset list
Hosting rights
Permission to store, package, and serve it.
A public license or written authorization must cover hosted inference for Windflow users.
Evidence · License URL or written permission · use restrictions
Runnable path
One clean launch path from input to output.
A Linux GPU environment, container, or launch command must run from a clean setup with one fixed reference input.
Evidence · Repository · setup · launch command · sample input
Measured baseline
A frame-rate number we can reproduce.
Name the hardware, resolution, precision, steps, batch, warmup, sample duration, and whether FPS is native or interpolated.
Evidence · GPU type and count · settings · FPS · method
Working terms
Clear before compute starts.
These are the defaults. Anything beyond them is agreed in writing before GPU work begins.
Subsidized integration
For selected open models, Windflow does not charge the model team for integration and covers public catalog serving.
Scoped per project
Dedicated capacity, vendor API integration, and joint projects are scoped and priced in writing before GPU work starts, including revenue sharing for hosted licensing.
Non-exclusive
Your team keeps its rights and can host, license, or integrate the model elsewhere.
Co-launch
Your name and logo on the catalog entry, and a joint launch post when the runtime goes public.
Delist on request
You can request removal at any time. We stop new sessions and agree a wind-down for anything already active.
Review before publish
We review the hardware, settings, method, and measured results together before any performance number goes public.
04 / What we cannot integrate
We do not build around a missing foundation.
Announcement only
A paper, teaser, or placeholder repository is not an inference release.
Weights unavailable and no endpoint
We cannot fetch the complete checkpoint at a stable revision, and there is no inference endpoint we can evaluate instead.
Hosting rights unclear
The license does not cover storage and served inference, and there is no written exception.
Baseline does not reproduce
The reference output or FPS depends on missing settings, a private service, manual steps, or an undisclosed interpolation path.
05 / Start the conversation
Send us the basics. Skip the deck.
A plain email is enough to book the intro call. Tell us which path fits, and we come back with next steps.
Email handoffPlain text is enough
- 01Model and team
- 02Preferred path: licensed weights, vendor API, or joint project
- 03Weights and exact revision, or an inference endpoint
- 04License or written hosting permission
- 05Reference GPU, settings, and measured FPS