Pika creates ultra-fast version of MiniMax H3: generation steps compressed from 50 to 6, with even better image quality.

Beating AI News Flash: Pika releases H3 Flex, creating a turbocharged version of MiniMax H3. The team first fine-tuned H3 with self-built audio-video data, then through model distillation compressed the denoising steps required for video generation from 50 steps to 6 steps. Subsequently, they improved image quality through reward and preference training, and further optimized underlying inference performance.

According to the DesignArena text-to-video leaderboard published by Pika, H3 Flex ranks 7th overall and 1st among open-weight models, surpassing the original MiniMax H3 and H3Max. In official tests, generating a 15-second 768p video with sound requires as little as 7 seconds of core model computation at the fastest (excluding video decoding and service processing time).

Pika has publicly released the model adaptation weights and inference code, supporting local operation via ComfyUI and Diffusers, with native maximum 768p video generation. The advertised 4K requires Pika's super-resolution service to achieve, and audio track separation and accelerated inference services have not been made available together.