pika-labs/pika-h3-flex is a repository for an H3-family model that generates video with synchronized speech from text and reference inputs. No separate latest release version is provided, and the repository description states that it covers video and speech generation inference. The README describes six-step generation and up to twelve reference inputs, and distinguishes between local weight execution and a hosted API.
The model is meant to align speech with video generation and to adjust scene composition with reference images. According to the README, it can accept up to twelve reference images and produces videos with synchronized speech attached through six-step sampling. Local execution natively supports 480p and 768p resolutions. The hosted API differs from local execution in resolution and features, but the provided facts do not spell out the specific differences. The repository's implementation language is Python, and it includes adapter, conditioning, sampling, geometry, and media handling modules.
The scope of the evidence extends to the generation path and the feature description. The README notes that the public demo material is not a performance benchmark, and it presents no reproducible throughput benchmark or validated results. The license is marked NOASSERTION, so usage conditions cannot be judged from this information alone. It is therefore difficult to evaluate actual performance or the range of practical use from the demo alone.