PikaStream1.0 (real-time video chat for agents)
PikaStream1.0 gives AI agents a live face with 24 fps 480p speech-to-video on one H100 at about 1.5 s latency.
A 9B diffusion transformer distilled into a causal streaming student, plus FlashVAE, a Transformer VAE decoding at 441 fps on 1.1 GB. Trained with multi-reward RLHF for identity, lip-sync and motion; agents keep memory and can act during the call.
- Date
- Thursday, 2 April 2026
- Lab
- Pika
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| End-to-end speech-to-video latency | ~1.5 s 24 fps, 480p, single H100 | company |
Pika's technical claims are self-reported.
Sources
This record was checked against its sources on 6 October 2026. How we check