Pika releases open-weight real-time video model
Pika has released open weights for H3 Flex, a video-and-audio model post-trained on MiniMax H3 that generates faster than real time and outputs up to 4K at $0.01 per second.
Pika has released H3 Flex, an open-weight "omni" video model family. Pika's team post-trained it on MiniMax H3, combining fine-tuning with reward-based optimization on an internal dataset chosen for aesthetic quality. Weights are on Hugging Face and code is on GitHub, and the model is also available through Pika's API, ComfyUI and Diffusers.
Pika says H3 Flex ranks first among open-source text-to-video models. It also claims the model beats leading competitors by "up to 91%" on visual quality, audio quality and human-preference evaluations. Pika ran all of these evaluations itself. In its human study, the model placed first for human anatomy and for artistic and aesthetic styles. The post does not show it winning every category.
The model outputs video from 480p to 4K. Pika describes it as real-time, meaning it takes less time to generate a clip than the clip lasts. H3 Flex also produces five separate, editable audio stems (dialogue, score, SFX, foley and ambience), which Pika calls a first for a video model. Other supported tasks include style and motion transfer, object replacement, character swaps, targeted edits, multi-shot generation, and generation from several image, audio or video references.
According to Pika's pricing tables, text- and image-to-video starts at $0.01 per output second. That matches HeyGen's tiers and is below the $0.08 listed for the MiniMax H3 base model, $0.103 for Seedance 2.5 and $0.101 for Gemini Omni 1.1 Flash. Reference-to-video starts at $0.03 per second.
Pika also offers custom training services.