Pika H3 Flex Generates 15-Second AI Video in 7 Seconds

The caveats

The 7-second figure refers only to the video-generating transformer. It excludes request handling, input preparation, decoding, audio processing and file delivery, so real waiting times will be longer, especially with reference files or higher resolutions. Clips run from 5 to 15 seconds at 24 frames per second.

The 4K output is also not native. The model generates natively at 480p and 768p only, while 1080p, 2K and 4K are produced by upscaling, which cannot restore detail that was never generated.

Editing and references

H3 Flex accepts up to 12 reference files at once: up to 9 images, 3 videos and 3 audio clips. Users can keep a product recognizable while changing the setting, lighting and camera movement, or swap a character or prop while keeping the original choreography. Multi-shot sequences with cuts, dialogue, music and effects are supported, and exports can include five separate audio tracks: dialogue, music, effects, action sounds and ambience.

Open weights, with limits

Pika has published the model's weights and code on GitHub and Hugging Face. The code uses the Apache 2.0 license, while the weights fall under the separate MiniMax H3 community license. The local version supports native 480p and 768p, but 4K upscaling, separate audio-track export and some tools remain exclusive to Pika's web service, and local speed depends on hardware.

The model is available through Pika API Club, with arrival in the main Pika tool announced for 2 November. Pricing is per second of video: 768p costs $0.04 per second at the regular rate, or $0.60 for 15 seconds, and launch pricing halves that. 4K output costs $0.16 per second, or $2.40 for a full clip.