Pika H3 Flex Generates Video With Separate Audio Stems, Starting at a Penny a Second
Pika H3 Flex generates 15-second clips with separate dialogue, score and SFX stems, from $0.01 per second. The weights are public, but the license excludes the U.S.
Pika has released H3 Flex, a new AI video model built on MiniMax H3 that generates clips up to 15 seconds long with a soundtrack. The soundtrack can be exported as separate audio stems. Pika says the model ranks first among open-weight models on DesignArena’s text-to-video leaderboard and seventh overall, ahead of the MiniMax models it was built from.
The model is available through Pika’s developer API at launch prices starting at $0.01 per second, and Pika has published the weights on Hugging Face. Pika says it’s “built for professional iteration,” with a focus on speed and on outputs that are easier to edit.
The Audio Stems
The most practical feature for editors is audio. H3 Flex generates native stereo sound with each clip, and Pika’s API can split that soundtrack into up to five WAV stems: dialogue, score, sound effects, Foley and ambience.
Most AI video models with audio bake everything into one mixed track. Separate stems mean you can replace the music, clean up dialogue or rebalance effects in your editor without regenerating the clip. Stem export costs $0.001 per second through the API.
For more on AI-generated sound for video, see Mirelo SFX 2.0 Brings AI Sound Design to Video.
Speed, Resolution and Pricing
Pika says H3 Flex can produce a 15-second clip with a soundtrack in as little as seven seconds on its hosted service. Clips run 5 to 15 seconds at 24fps, from text, images or keyframes, and the model also supports reference-to-video.
Generation happens at 480p or 768p, and Pika’s hosted service offers 1080p, 2K and 4K finishing from a draft video. Launch pricing per second of output is $0.01 at 480p, $0.02 at 768p, $0.035 at 1080p, $0.04 at 2K and $0.08 at 4K, half of Pika’s regular rates.
H3 Flex is available through Pika’s developer platform, not the consumer Pika app.
The “Open Source” Fine Print
Pika describes H3 Flex as open source from day one, with code on GitHub, weights on Hugging Face and support in ComfyUI and Diffusers. The inference code is under the Apache 2.0 license, but the model weights are covered by the MiniMax H3 Community License, since H3 Flex is built on MiniMax H3.
That license excludes the United States, the European Union, the United Kingdom and South Korea from its grant. In practice, that means running the downloaded weights isn’t licensed for users in those regions, even though the files are publicly available.
The license also requires separate permission from MiniMax for products earning more than $20 million a year, and requires “MiniMax H3” to be displayed in the interface of commercial products. For U.S. creators, Pika’s hosted API is the straightforward way to use the model.
Running it locally is also demanding. Pika tested it on an H100 80GB GPU, the base components total about 210GB on disk, and Mac support isn’t confirmed.
Competitive Context
H3 Flex is priced well below most hosted video models. Kling 4.0 runs about $0.157 per second at 720p on fal, and Seedance 2.5 about $0.47 per second at 720p on fal, covered in How to Get Started With Seedance 2.5.
MiniMax’s own H3 Max, which H3 Flex outranks on DesignArena according to Pika, is also available through hosting platforms like fal. Leaderboard rankings are based on crowd preference votes, so side-by-side tests on your own prompts are a better guide than the ranking alone.
The Signal in the Noise
For creators, the stems and the price are the story. Getting dialogue, music and effects as separate files at a fraction of a cent per second makes AI clips much easier to drop into a real edit and fix there.
The “open source” framing deserves an asterisk. The weights are public, but the license doesn’t cover U.S. users, and the 4K output and fastest speeds come from Pika’s hosted service rather than a local install.
Would separate audio stems change how you use AI-generated clips in your edits? Share how you’d handle the soundtrack in the comments.
Specs & Pricing
- Model: Pika H3 Flex (post-trained on MiniMax H3)
- Clip length: 5–15 seconds, 24fps
- Generation resolution: 480p or 768p; 1080p, 2K and 4K finishing via Pika’s API
- Audio: Native 32kHz stereo; up to five WAV stems (dialogue, score, SFX, Foley, ambience)
- Launch pricing per second: $0.01 (480p), $0.02 (768p), $0.035 (1080p), $0.04 (2K), $0.08 (4K)
- Stem export: $0.001 per second
- Access: Pika developer API; weights on Hugging Face; ComfyUI and Diffusers support
- Weights license: MiniMax H3 Community License (excludes U.S., EU, UK and South Korea)
- Code license: Apache 2.0
