Pika releases H3 Flex, an open-weight video model with separate audio stems
Pika's new model is post-trained on MiniMax H3, renders up to 15 seconds with sound and splits the soundtrack into five editable tracks. Pika says it starts at $0.01 per second.
Pika has launched Pika H3 Flex, a video model it post-trained on MiniMax H3. In its technical write-up, Pika says the model makes clips of up to 15 seconds with synchronized stereo audio, and that it generates faster than the clip plays back.
The model renders natively at 480p and 768p. A separate upsampler takes the result to 1080p, 2K or 4K. Pika says H3 Flex is the first video model to hand you the soundtrack as five editable stems: dialogue, score, sound effects, foley and ambience.
The weights are on Hugging Face, and the model runs through Pika’s API, ComfyUI and Diffusers. Pika lists a starting price of $0.01 per output second. Pika also says H3 Flex beat MiniMax H3, fal’s H3 Max and HeyGen Video in its own human tests on anatomy, composition and audio sync. Those tests are Pika’s own and haven’t been checked independently.
What it means for you: if you mix your own audio, separate stems save you a round trip through a stem-splitter. Pika hasn’t said what H3 Flex costs in credits inside its consumer app.
Source: Pika blog, 29 Oct 2026
