Realtime Videos · Resources
FAQ
Is this production-ready?
No. It is a preview. The API, queueing and metering are tested against a CPU test model, and the real-time model has not yet been measured on production GPUs. Use it to build and evaluate, not to ship.
How is this different from the offline video API?
The offline API (/v1/videos) is queue-and-poll: submit a job, wait, download a file. Realtime opens a live session, streams frames as they are generated, and accepts commands while running. They are separate APIs for separate models. See the offline docs.
Can I use MiniMax-H3 in realtime?
No. It generates a whole clip per request and keeps no state between chunks, so it cannot be steered live. It is served by the offline API instead.
Can I use this from a browser?
Not safely yet. There is no browser SDK and no short-lived client token, so a browser would need your API key. Create sessions and proxy signaling from your own backend. See Authentication.
My session stays queued.
Every slot for that model is busy. Each pod runs one session at a time, so capacity is the number of pods. The session fails with capacity after 120 s by default; retry later or end idle sessions.
I get 404 "no realtime pods serve model".
No pod for that model is registered right now. Check the model name against the catalog and confirm with us that a pod is running.
The WebRTC connection never opens.
ICE could not find a path. Media needs UDP between you and the pod, and direct connectivity from rented GPUs is not yet verified. Test from the network you deploy on, and expect to need a TURN relay on restrictive networks.
My prompt change took a moment to show.
Commands apply at the next block boundary, and a prompt switch makes that block slower. For longlive-1.3b a block is 750 ms of video. This is how block-by-block generation works, not a bug.
Does a paused session cost less?
No. It keeps its GPU slot and the meter keeps running. End it instead.
Is prompt or output moderation applied?
Not in the preview. Do not expose realtime to untrusted end users until it is.