Pika has released Pika H3 Flex, a video generation model built on MiniMax’s H3, and published its adapters and inference code for anyone to download. The company announced it on its blog on 29 October. The pitch is speed for people who regenerate the same shot a dozen times before they keep one.

The speed comes from cutting the number of denoising steps. Pika says it retrained H3 so a video takes six denoising passes where the original needed 50, then tuned the faster version for quality. The headline figure is seven seconds of transformer computation, the best case for a 15-second clip with sound at 768p. Pika’s own footnote narrows that number: it comes from a controlled benchmark of the transformer alone, and the company says reference inputs, upscaling, audio separation, hardware and service overhead all add to the time you actually wait.

On quality, Pika cites the DesignArena leaderboard for text-to-video, where it says Flex took first place among open-weights models and seventh place overall. That is a launch-day snapshot of a third-party ranking, and Pika says as much.

The feature that separates Flex from a plain speed-up is how much you can feed it. A single generation takes up to 12 reference files, mixed freely across images, video and audio. In Pika’s examples a hero comes from one picture, the setting from another, and the choreography from a clip. A sketch, a product photo, a moodboard or a voice sample all count as references, and the user tells the model what each one should contribute. For a studio making ads, that means a product shot can be set in new locations without the product drifting from its photographed look. Pika’s lip-gloss commercial demo does exactly that from one product image.

Editing works the same way. Given original footage, Flex can change clothing, props or surroundings while keeping the performance and timing of the source clip. Shot direction can be specific too: Pika’s examples stack an overhead opening, a crane move, a quick pan and a circling shot into a single clip.

Flex produces stereo sound in sync with the picture, and through the Pika API it can also return five separate stems: dialogue, score, sound effects, Foley and ambience. An editor who wants the music lower without touching the conversation can pull the tracks apart instead of regenerating the clip. Stem separation is one of the features Pika hosts itself, so it will not be in the downloadable package.

Resolution follows the same draft-then-finish logic. The model itself renders at the two lower sizes, 480p and 768p. A separate upsampler from Pika then lifts a chosen clip to 1080p, 2K or 4K, which the company says preserves the generated content while adding detail. The upsampler is also hosted by Pika.

The open release includes text-to-video, generation from a first and last frame, and the multi-reference mode, with sound at the two lower resolutions. Weights sit on Hugging Face and code on GitHub, governed by a license whose terms the blog post does not spell out. Pika also notes that local performance depends on hardware, and it gives no numbers for what a single consumer GPU would manage.

Commercial access starts through the Pika API Club, and Flex arrives in Pika’s creation product on 2 November.

The draft model is open, while 4K finishing and stem separation stay behind Pika’s API. A team that wants its own pipeline can start testing the 768p model this week, and everyone else should price the hosted finishing step before assuming an open model means a free workflow.

Reported by the Pika blog (Pika Team) on 29 October 2026.