What happened
The downloadable release generates video with synchronized audio at 480p or 768p. It supports text prompts, first- and last-frame conditioning, and multiple references. Higher-resolution finishing to 1080p, 2K and 4K, audio-stem separation, prompt enhancement and accelerated serving remain hosted features.
Pika reports seven seconds of transformer computation for a 15-second, 768p clip in a controlled test. This excludes complete API processing time. Its technical report describes six sampling steps, compared with the teacher model’s 50.
The model card identifies the adapters as MiniMax H3 derivatives under its Community License. Representative local tests used an H100 with 80GB of memory; performance depends on hardware and configuration.
Why it matters
For production teams, the development creates a choice between operating the generator themselves and buying the hosted finishing workflow. Those options need separate cost and quality tests. Downloadable weights do not establish that local performance matches the hosted service.
What to watch
Measure finished clips accepted by an editor, reference consistency and complete turnaround time. A short computation benchmark alone cannot show the cost or reliability of an entire production process.
Sources & verification
Source links are not a count of independent confirmations.
01Report a correction
Help us get it right
Include the disputed statement and, if possible, a supporting source URL. Please do not send personal or sensitive information.
-00.06.31.698.png)
