Pika-Labs releases Pika-H3-Flex open weights on Hugging Face
Pika Labs has published Pika-H3-Flex on Hugging Face as an open-weight image-to-video model, giving creators and researchers direct access to weights they can run locally. The release arrived on October 29, 2026, and has accumulated nearly 500 likes on the platform in short order, signaling notable interest from the generative video community.
The model is built on a post-trained H3 architecture and is designed with production workflows in mind. Its most notable technical characteristic is a six-step sampling process, which keeps generation fast without sacrificing control. It also supports what Pika calls "Omni-reference" - accepting up to 12 reference inputs simultaneously, allowing shot direction to be guided by a combination of images and text prompts. The locally runnable weights produce native 480p and 768p video with mixed stereo audio baked in, making it one of the few open-weight models to handle audio as part of the core generation pipeline rather than as a separate step.
For teams that need higher output resolutions, Pika's hosted API extends the same model to 1080p, 2K, and 4K, and adds features like isolated audio stems and prompt enhancement. This two-tier approach - open weights for experimentation and a commercial API for production-grade output - reflects a strategy several model developers have adopted to balance community openness with sustainable business models. The model card is explicit that the creative showcases shown are not throughput benchmarks, and exact reproduction of those samples with the released package is not guaranteed.
Pika-H3-Flex appears aimed at professional use cases including multi-shot narrative work, motion graphics, product films, and reference-guided animation. The emphasis on camera choreography and cinematic control through reference inputs positions it as a tool for creators who need repeatable, directed output rather than purely generative exploration. It is described as one member of a broader H3 video-and-audio model family, suggesting further releases in this line may follow.
