Skip to content
File

Pika: video and audio generation models

This dossier follows Pika’s video and audio models, distribution terms and the distinction between local capabilities and hosted services.

Last updated : 1 key moments

Where things stand

As of 30 October 2026, H3 Flex adapters and code are available. Local generation is limited to 480p or 768p; 4K finishing and separate audio stems use Pika’s hosted services.

Pika releases H3 Flex: local video and audio, hosted 4K finishing

Pika released H3 Flex adapters and inference code on 29 October 2026. Derived from MiniMax H3, the model generates video with synchronized audio locally at 480p or 768p.

The documentation supports up to twelve references combining images, video and audio. Upscaling to 1080p, 2K or 4K and exporting five separate audio stems are Pika-hosted features, absent from the local package. Adapters use the MiniMax H3 Community License; the code’s Apache 2.0 license does not cover weights.

Pika advertises fifteen seconds of video computed in seven seconds, but that measures only a computation stage, not full service latency. Availability in its creation app is announced for 2 November.

Sources 4