Pika-Labs/Pika-H3-Flex: Professional iteration video and audio model on Hugging Face

Pika Labs introduced Pika H3 Flex, a post-trained H3 video-and-audio model family designed for professional creators, featuring two task adapters for text and multimodal reference conditioning. The model generates native 480p and 768p video with synchronized stereo audio in six sampling steps, while a hosted API extends capabilities to higher resolutions including 4K. Built on MiniMax-H3 base components, the inference package is available for local open-weight use and integration with ComfyUI and Hugging Face Diffusers. Open article →