Pika introduced H3 Flex on Oct. 29 as an open-weight video model. The company says it post-trained the model on MiniMax H3. According to ComfyUI’s guide, H3 Flex runs as two LoRA adapters on the MiniMax H3 base, rather than as a new model architecture.
What seven seconds measures
Pika’s launch post advertises a 15-second clip with a soundtrack in as little as seven seconds. Its FAQ narrows that figure to transformer computation for a 15-second, 768p clip with synchronized audio in a controlled benchmark. Pika explicitly says it is not total end-to-end API latency: higher-resolution finishing, audio-stem separation and service overhead can add time to a request.
What comes with the open weights
Pika says the public release includes generator adapters and inference code for text-to-video, first-and-last-frame generation and multimodal reference-to-video at 480p and 768p. It links to GitHub code and Hugging Face weights; ComfyUI provides workflows for text prompts, reference images and animation between two still frames.
The higher-resolution tools have a different availability story. Pika says native generation is at 480p or 768p, while its hosted upsampler finishes clips at up to 4K. Its API can also return separate dialogue, score, sound-effects, foley and ambience tracks. Neither the upsampler nor audio-stem separation is part of the open-weight release, according to Pika.
Pika claimed at launch that H3 Flex ranked first among open-weight models on DesignArena’s text-to-video board. Pika also claimed at launch that H3 Flex ranked seventh for video quality on DesignArena’s text-to-video board. For anyone running the public adapters, ComfyUI notes that they are distributed under the MiniMax H3 Community License Agreement and advises reviewing its territorial and commercial-use conditions.
Comments