Realtime Videos
Build with live, interactive video models
Open a session, watch frames arrive over WebRTC as the model generates them, and change the prompt while it runs.
16 fps
Playout rate, LongLive-1.3B
750 ms
Per block: when a command can take effect
1 s
Billing granularity, from ready
What the realtime API is
Most video APIs are queue-and-poll: you submit a job, wait a minute or two, and download a file. The realtime API is different. It is session-oriented: you ask for a session on a GPU, connect to it over WebRTC, and receive video in short blocks as the model produces them. You can send commands, such as a new prompt, while the stream is running and see the result in the next block.
- HTTPS control plane. Create, inspect and end sessions with your
tf_API key. - WebRTC media plane. Frames and live commands go peer-to-peer between your client and the GPU. They do not pass through our API servers.
- Typed commands. Each model publishes its own command schema, so new models need no API change.
- Per-second billing while a session holds a GPU. Waiting in the queue is free.
How a session flows
| Step | You | We |
|---|---|---|
| 1 | POST /v1/realtime/sessions | Pick a free GPU pod for the model, reserve its slot, start the session. |
| 2 | Poll until status is ready | If no slot is free the session is queued, with a position and an estimate when one exists. |
| 3 | Negotiate WebRTC through the signaling proxy | Authenticate and forward signaling to the pod. Media never touches us. |
| 4 | Receive the video track, send commands | The pod generates a block, paces it to the playout rate, and applies commands at the next block boundary. |
| 5 | DELETE the session | Free the GPU and report final usage. |
Availability
| Capability | State | Notes |
|---|---|---|
| Sessions API, queue, metering | Preview | Tested against a test model; not yet run against a production model on a GPU. |
| Signaling proxy (WebRTC) | Preview | Verified on loopback. Reachability from rented GPUs (UDP, TURN) is not yet verified. |
| Client SDKs | Not yet | A reference Python viewer exists. There is no published browser SDK. |
| Content moderation | Not yet | Neither prompt nor output moderation is enforced in the preview. |
| Published pricing and SLA | Not yet | Usage is metered in seconds; rates are not published. |