Pika H3 Flex: the open-weight video model with audio for creators who redo the same shot ten times
Pika launched Pika H3 Flex on October 29, 2026: a video model that generates 5 to 15-second clips with synchronized stereo audio, accepts up to 12 references and can deliver up to five separate audio stems. The weights are public on Hugging Face, but under a license worth reading. Here is what it actually does, what it costs per second and what it changes for AI video creators.
Leggi questo articolo in italiano
What Pika H3 Flex is and when it launched
Pika H3 Flex is an AI model that generates video with audio, launched by Pika on October 29, 2026 with a press release from San Francisco and an official page with specs and pricing. It is not built from scratch: it is a version of MiniMax H3, MiniMax's open video-and-audio model, post-trained by Pika's research team.
The thinking behind the name "Flex" is spelled out in the press release: most video models assume your first generation is the one you keep, while people working for clients redo the same shot many times. So Pika focused on three things: speed, a low cost per attempt and the ability to fix a video without starting over.
The core technical change is fewer generation steps: 6 instead of 50. According to Pika, a 15-second 768p clip with audio takes as little as 7 seconds of model compute. Note the fine print: the same official page says this is transformer (DiT) time only, measured in a controlled benchmark, and not total request time, which grows with references, upscaling and audio stems. Every speed and cost figure in this article is claimed by Pika.
What it does: 12 references, editing and 5 separate audio stems
Pika H3 Flex generates clips from 5 to 15 seconds at 24 frames per second, in 11 aspect ratios from 21:9 to 9:21, with prompts up to 8,000 characters. Here are the modes and limits listed in the official documentation:
| Feature | What it does |
|---|---|
| Text to video | Clips with synchronized 32 kHz stereo audio |
| First and last frame | Starts from an image, ends on an image, or both |
| Reference to video | Up to 12 references: 9 images, 3 videos and 3 audio clips; at least one image or video is required |
| Guided editing | Swaps an outfit or a prop while keeping the original performance and timing |
| Separate audio stems | Up to 5 WAV files: dialogue, score, SFX, Foley and ambience |
| Draft and finish | Draft at 480p or 768p, then take the same shot to 1080p, 2K or 4K |
Video and audio references must each run between 2 and 15 seconds, with a 15-second combined cap per type. You cannot start from audio alone: at least one image or video is mandatory.
For editors, the most practical feature is the separate audio stems. Instead of a single mix, you can request dialogue, score and effects as distinct files and lower or replace the music in your edit. The model generates natively at 480p or 768p: higher resolutions come from an upscale in the same job, and with draft-to-video you only pay the difference between the draft and the final version.
How much Pika H3 Flex costs per second
Pika H3 Flex is billed per second of generated video through Pika API Club. These are the prices published on the official page for text, image and keyframe generation:
| Resolution | List price | Launch price |
|---|---|---|
| 480p | $0.02/s | $0.01/s |
| 768p | $0.04/s | $0.02/s |
| 1080p | $0.07/s | $0.035/s |
| 2K | $0.08/s | $0.04/s |
| 4K | $0.16/s | $0.08/s |
In practice, at list price a 15-second clip costs about $0.30 at 480p and $2.40 in 4K. Reference-to-video costs more (from $0.06/s at 480p list price) and also bills the seconds of uploaded reference videos, while image and audio references are free. Separate audio stems cost $0.001 per second.
Pika says H3 Flex is up to 1.8 times faster and up to 80% cheaper than H3 Max. The official page lists 15 free generations a day for API Club members from October 29 to November 5, and VP Land confirms the entry price of $0.01 per second. The press release adds that the model reaches Pika's creation app on November 2.
Open weights: what you can download and what you cannot
Pika published the weights and inference code on launch day, on the Pika-Labs/Pika-H3-Flex Hugging Face page and on GitHub, with ready-made ComfyUI workflows and Diffusers support. Hugging Face lists 33 billion parameters.
There are three limits to know before getting excited:
- The license is not as permissive as Apache or MIT. The weights are derivatives of MiniMax H3 and fall under the MiniMax H3 Community License Agreement; only the example code is Apache 2.0. Read the license before any commercial use.
- The open version stops at 480p and 768p. Upscaling to 1080p, 2K and 4K, audio-stem separation, prompt enhancement and accelerated serving stay on Pika's platform.
- It needs data-center hardware. The tests listed on the model page ran on an 80 GB H100 GPU, and the base components take about 210 GB of disk space.
On quality, Pika says H3 Flex ranks seventh for video quality on DesignArena's text-to-video board, with a score of 1,293, ahead of MiniMax H3, H3 Max and Seedance 2.5, and that it is the top model with open weights. Pika itself notes that its showcase videos are not official benchmarks.
What it means for you as an AI video creator
Pika H3 Flex makes getting it wrong cheaper, and that is the most expensive part of the job. Here is how to use it:
- Work in two passes. Run your tests at 480p, where 15 seconds cost a few cents, and take only the right shot to 1080p or 4K with draft-and-finish.
- Prepare references before the prompt. With 9 images, 3 videos and 3 audio clips you can lock character, product, location, camera move and voice: the safest way to keep clips consistent.
- Ask for stems if you edit. Dialogue, score and effects as separate files let you change the music or clean up the voice without regenerating the video.
- Fix instead of redoing. Reference-guided editing swaps an outfit or a prop while keeping the performance, which is handy for variants of the same ad.
- Do not download the weights out of habit. Without an 80 GB GPU and a careful read of the license, the pay-as-you-go service is the practical route for a creator.
To choose between today's video models, our guide Kling vs Seedance vs Veo (in Italian) compares their strengths, and here is what changed with Kling 4.0. We saw the same return of open models on the text side with Liquid AI LFM3. To write prompts that make the most of shots and camera moves, see our guide to cinematic AI video prompts. If you want to turn models like this into ads, trailers and short films with a method, the AI Filmaking program takes you from concept to final cut; to edit clips and stems fast there is AI Editing Automation; and if you would rather receive a finished video, our custom 30-second AI video service does it for you. Here is also how we use artificial intelligence.
Frequently asked questions
When was Pika H3 Flex released?
Pika announced Pika H3 Flex on October 29, 2026, publishing the weights and code the same day. According to Pika's press release, it reaches the Pika creation app on November 2, 2026.
Is Pika H3 Flex open source?
The weights can be downloaded from Hugging Face, but the license is the MiniMax H3 Community License, not a permissive license like Apache or MIT. The open version generates at 480p and 768p; upscaling and separate audio stems stay on Pika's service.
How much does Pika H3 Flex cost?
List pricing starts at $0.02 per second at 480p and goes up to $0.16 per second in 4K; launch prices are half that. Reference-to-video with video references costs more.
How long can a Pika H3 Flex video be?
From 5 to 15 seconds per generation, at 24 frames per second, with synchronized stereo audio.
What are Pika H3 Flex audio stems?
Through the API, Pika can deliver up to five separate WAV files: dialogue, score, sound effects, Foley and ambience, so you can remix the audio in your edit.
Read next



Sources checked on October 30, 2026
- Pika — Pika H3 Flex: faster-than-realtime video model (pagina ufficiale con prezzi e specifiche, 29 ottobre 2026)
- Pika Labs su Hugging Face — scheda del modello Pika-H3-Flex (pesi, licenza, requisiti)
- VP Land — Pika H3 Flex Is a New Open-Weight Video Model Built for Fast Iteration (29 ottobre 2026)
- Pika (Business Wire, su Yahoo Finance) — Pika Launches Pika H3 Flex, a Video Model Built for Professional Iteration (29 ottobre 2026)