Pika releases H3 Flex as an open-weight video model

Pika has released Pika H3 Flex, a text-to-video model the company debuted as the number 1 open weight model on DesignArena's t2v leaderboard. Open weight means the underlying model parameters are published for anyone to download and run themselves, rather than only being accessible through Pika's own service. Pika says the model is built for professional iteration, with exportable audio stems, meaning the generated soundtrack can be pulled out as a separate audio track, plus resolutions up to 4K.

According to Pika, the model generates faster than realtime, producing a 15-second clip with a soundtrack in as little as 7 seconds. In a technical follow-up post, the company said it post-trained H3 Flex to generate video in six denoising steps instead of 50. Denoising steps are the iterative passes a model takes to turn random noise into a finished image or video, so cutting that count sharply is a major source of the speedup; Pika said it then refined the model further for quality and creative control. Pika says pricing starts at 0.01 per second.

Why it matters

An open-weight video model that Pika claims generates faster than realtime could make quick video iteration far cheaper for developers to self-host.

Source: Pika on X