PikaStream 1.0
#54 en Texte-vers-vidéoPika Labs · v1.0 · depuis 2. April 2026 (Beta) · 10× · vu le 30 juin 2026
PikaStream 1.0 is a real-time video engine model from Pika Labs that powers the "pikastream-video-meeting" skill, giving any AI agent a face, voice, and the ability to join live Google Meet video calls and execute tasks mid-conversation. Technically, it is built on a 9-billion-parameter Diffusion Transformer plus a component called FlashVAE, generating personalized video at 24 FPS at 480p resolution on a single H100 GPU with an end-to-end speech-to-video latency of about 1.5 seconds. The model was officially announced on April 2, 2026 as a beta release, distributed via the open-source Pika Skills repository on GitHub and the Pika Developer API. Official first-party pricing is not clearly stated on Pika's own site; third-party sources cite conflicting figures ranging from $0.20 to $0.50 per
Fonctionnalités
| Fine-Tuning | No classical model fine-tuning; personalization is via voice cloning from a short audio sample and a custom/generated avatar image |
| Generation Time | 24 FPS streaming frame rate, ~1.5 seconds end-to-end speech-to-video latency on a single H100 GPU |
| License | The pikastream-video-meeting skill is open source on GitHub (Pika-Labs/Pika-Skills); the PikaStream1.0 model itself is proprietary, accessed via API |
| Max Resolution | 480p during real-time streaming |
| Platform | Google Meet (Zoom and FaceTime planned); access via Pika app, Pika Developer API, and open-source GitHub skill |
| Price | Inconsistently reported: $0.20–$0.50 per minute of usage (official Pika pricing page does not clearly state a figure) |
| Release Date | April 2, 2026 (beta release) |