Pika Labs Releases Open Weight H3 Flex Video Model

Pika Labs released Pika H3 Flex, an open weight video generation model, as detailed on its Hugging Face model page. The model generates synchronized audio and video in six sampling steps and accepts up to 12 reference inputs.

Local inference supports native 480p and 768p video with mixed stereo audio. Users can direct shots through prompts and reference material for tasks including multiple shot sequences, motion graphics, product films, and reference guided animation.

The hosted Pika API adds generation at 1080p, 2K, and 4K resolutions. It also provides separate audio stems and prompt enhancement.

The 33-billion-parameter release builds on MiniMax's MiniMax-H3 model and ships as LoRA adapters for text, endpoint-frame and multimodal-reference tasks. It is licensed under the MiniMax-H3 community license agreement.