Pika releases open-weight video model with synchronized audio
H3 Flex runs locally as a six-step LoRA on MiniMax H3, while Pika offers hosted 4K upscaling and separate audio stems.
Pika has released the weights and inference code for H3 Flex, a video-and-audio model built by post-training MiniMax H3. The system generates clips with synchronized stereo audio in six sampling steps and supports text prompts, first and last frames, and multimodal references.
Local inference produces 5 to 15-second clips at native 480p or 768p across 11 aspect ratios. Pika packages two LoRA adapters: one for text and endpoint-frame conditioning, and another for image, video and audio references. The company says a selected workflow has been tested on an H100 GPU with 80 GB of memory; the complete base components occupy about 210 GB, though only the relevant partition loads for a request.
Pika advertises output up to 4K, exportable audio stems and a 15-second clip in as little as seven seconds. Those features belong to its hosted API, not the downloadable local package, and the speed claim has not been independently validated.
On DesignArena, H3 Flex debuted seventh overall with a 1293 Elo score and was the highest-ranked open-weight entry on the text-to-video board. Arena rankings can move as more votes arrive, so they are a useful signal rather than a definitive quality measure.
The release is notable because Pika had previously kept its main generation models closed. The adapters inherit the MiniMax H3 Community License Agreement, so open-weight does not mean OSI open source. Developers can nevertheless inspect, adapt and run the model instead of relying only on a hosted generation service.
