Pika just gave away a video AI that renders faster than real time

Pika has released its new video model, H3 Flex, with open weights. It came out on the evening of October 29, and it is a rare case where a serious video generator with sound is not just something to try through a service but something you can download whole.

Under the hood is China's MiniMax H3, which Pika post-trained for speed: a clip now takes six generation steps instead of fifty. By the company's own benchmark, a 15-second clip with synced audio at 768p takes about seven seconds of model compute. The fine print is honest: that counts only the neural network itself, not queueing, reference loading or upscaling, so real waits are longer.

The headline feature is references. One generation can take up to 12 files at once: images, video and audio. The character comes from one picture, the world from another, the movement from a clip of someone dancing, and it all lands in one scene. You can also take finished footage and ask it to swap the outfit, a prop or the background while keeping the performance and timing. Through the API, sound comes back as five separate stems: dialogue, score, effects, foley and ambience, so you can turn the music down without touching the dialogue.

The model natively outputs 480p and 768p, and clips can be upscaled to 1080p, 2K or 4K, but upscaling, stem separation and the fast serving stack stay on Pika's side. What is open are the adapters and inference code on Hugging Face and GitHub. At launch, Flex ranked first among open-weight models and seventh overall on DesignArena's text-to-video leaderboard.

Running it at home is not for everyone: the base files take around 211 GB, you need an NVIDIA GPU, and it was tested on an H100 with 80 GB of memory. Everyone else can try it through the Pika API Club, and on November 2 it lands in the regular Pika app.

Verdict: video with sound from a dozen references, free and with open weights. Heavy for a home PC, a gift for studios and developers.