Pika H3 Flex: an open AI video model built for fast drafts
Pika H3 Flex makes 5–15-second AI videos with sound from up to 12 references, drafts fast at 768p, and its weights are open. What it means for brands.
By Olena, me20 Studio · 3 min read

Pika H3 Flex is a new AI video model from Pika, launched on 29 October 2026. It turns text, keyframes or up to 12 reference files into 5–15-second clips with synchronised sound, drafts quickly at 768p and finishes up to 4K. Unusually for a top video model, its weights are open, so studios can also run it on their own hardware.
What is Pika H3 Flex?
H3 Flex is Pika's own "post-trained" version of the MiniMax H3 video model, tuned for people who revise a lot. Pika cut generation from 50 steps to six and says a 15-second 768p clip with audio can take as little as seven seconds of model compute — a lab benchmark, not the full waiting time you will see.
At launch it ranked first among open-weights models and seventh overall on DesignArena's text-to-video leaderboard (Pika's figures, a snapshot that will change).
Twelve references and sound in one go
One generation can take up to 12 references: up to 9 images, 3 videos and 3 audio clips. Pika lists sketches, character sheets, product photos, moodboards, voice samples and movement videos as examples; at least one image or video is required.
- clips of 5 to 15 seconds at 24 fps, 11 aspect ratios from 21:9 to 9:21
- native 480p or 768p, with a hosted upscaler to 1080p, 2K or 4K
- synchronised stereo audio; through the API you can export five stems — dialogue, score, SFX, Foley and ambience
- reference-guided edits: change an outfit, prop or setting while keeping the performance and timing
Open weights: what is actually open
Pika published the model's adapters and inference code on Hugging Face and GitHub, with ComfyUI workflows. The open version does text-to-video, first-and-last-frame video and reference-to-video with audio at 480p and 768p. The 4K upscaler, audio stems and Pika's fast serving stay on Pika's own service.
The adapters sit on top of the MiniMax H3 base model (about 210 GB of files) and are covered by the MiniMax H3 Community License — read it before any commercial use. Pika's own tests ran on an 80 GB H100 GPU.
Price and how to try it
H3 Flex is available now through the Pika API (dev.pika.art) and, according to Pika, arrives in the Pika app on 2 November. Regular prices start at $0.02 per output second at 480p and $0.04 at 768p; until 5 November launch prices are half that, and Pika API Club members get 15 free generations a day.
A practical way to test it: draft at 768p, compare three or four versions of the same shot, and pay for 1080p or 4K only for the one you keep. Write the camera move into the prompt and export separate stems if you plan to remix the sound.
Writing your first prompts? See our camera moves glossary, the guide to sound design for AI video ads and our video work. Want a short film or ad made with the right tools for your brand? Tell us about your project.
P.S. The caption under the cover image is a ready-to-use prompt. Copy it into Midjourney, Higgsfield or any image generator and make your own version.