Pika releases H3 Flex, an open video model

Pika's post on 29 October 2026 introduces Pika H3 Flex, a post-trained MiniMax H3 model. Pika says it generates in six steps instead of 50, accepts up to 12 references that can mix images, video, and audio, and can export five audio stems: Dialogue, Score, SFX, Foley, and Ambience. Native output is 480p or 768p with synchronized audio. A hosted upsampler finishes a selected clip at 1080p, 2K, or 4K.

Pika says a 15-second clip at 768p with synchronized audio can take as little as seven seconds of transformer computation. The same post says that figure is compute time in a controlled benchmark, not end-to-end API latency. At the time of the post, Pika places Flex first among open-weight models and seventh overall on DesignArena's text-to-video board, and says that rank is a launch snapshot.

The open release is the generator adapters and inference code, for text-to-video, first- and last-frame generation, and multimodal reference-to-video, with synchronized audio at 480p and 768p. Pika says the upsampler, stem separation, optional prompt enhancement, and accelerated serving stay on its hosted stack. Flex is available through the Pika API. Pika says it comes to the Pika creation product on 2 November 2026.

Who it's for: someone iterating on a short clip through the API, or a developer running the open weights.

Fits: Best for Video stays Runway. Pika stays fifth, for a social clip, an effect, or a joke. Flex is a faster model with more references. It is not the editing workflow the first rank is for. The video guide, the Pika alternatives, and the Runway alternatives stay on that verdict. The verdict is unchanged.