Pika H3 Flex: open video weights, the finish stays in the cloud
Pika just dropped open weights for a video model that eats up to twelve references and brings the sound along. What that means for the way we work, and what still lives on Pika's servers.
Miraki MindAI-assisted editorial · 30 October 2026 · 6 min read
On 29 October 2026, Pika released Pika H3 Flex, a video model built on MiniMax H3 and post-trained by Pika. Pika's adapters and the code are open on Hugging Face and GitHub. The thing makes picture and stereo sound in one go: from text, from a first or last frame, or from a whole stack of references.
Every film we make starts with references. Characters, places and props are built as sheets first and approved before a single shot is made. An example from It Ain't Coming Off: behind the film still below are sheets like the location sheet of the desert and the character sheet of Dee. The shots themselves were then made with Seedance 2.5 in the cloud.
So an open model that understands exactly these kinds of references, runs locally and brings the sound along hits two questions we deal with all the time: how fast do we get from an approved sheet to the first moving image? And where does a brand's material end up while we work with it?
What is actually new
Open video models have been around for a while. What's new is how much control an open model now gives you. And that the sound comes with it.
- References: up to twelve images, videos or audio files per generation. In the prompt, you say what each reference should contribute.
- Editing: according to Pika, a reference can swap out part of a shot, like an outfit, a prop or a location, while the performance and timing stay as they were.
- Sound: stereo, in sync with the picture. Through Pika's API you can also get it as separate tracks: dialogue, music, effects, foley and ambience.
- Speed: Pika has trimmed the model down to six denoising steps instead of fifty. The speed figure Pika quotes comes from a controlled benchmark, though, and according to the footnote it is not the time a request takes end to end.
The model card is just as honest about the limits: native 480p or 768p, 24 frames per second, five to fifteen seconds. The upscale to 1080p, 2K or 4K, the split into tracks and the fast serving only exist on Pika's own platform.
What it means for a production
Twelve references with a clear role each: that's pretty much our workflow. The approved character, the location and the prop in the same generation. And if a prop changes, Pika says the whole shot shouldn't have to be redone.
For final shots, though, the open part alone won't cut it. 768p is too little for a commercial, and the step to 4K runs on Pika's servers. For delivery, there's no way around the cloud models for us right now.
Where Flex gets interesting is earlier in the process: look tests, previz, animatics, lots of variations in little time, straight from our sheets. Just not on a laptop. Pika mentions test runs on an H100 with 80 GB, and the base components need about 210 GB of disk space. The model card links ComfyUI workflows and Diffusers examples, so you can plug it into an existing node pipeline.
What it means for a brand
Running it locally means the references stay on your own hardware. A product that isn't out yet, faces from the cast, campaign material before launch. For data protection, that's a real difference, but only up to the finish. If a shot needs the hosted upscale, the material leaves the building anyway.
The licence is worth a close read. The adapters come under the MiniMax H3 Community License Agreement, and the bundled base components keep that agreement. Pika's announcement doesn't summarise the terms. Open weights means the files are available. What you're allowed to do with them in client work is up to the licence.
Labelling doesn't change. AI-generated footage gets labelled, no matter which model it comes from.
A model that takes a dozen references fits a way of working where the look gets approved first.
Our take and the limits
Pika H3 Flex is a serious tool for the early phase of a film, where speed and control over the references matter more than resolution. For delivery it doesn't replace the cloud models, and Pika itself keeps the finish on its own platform. The ranking Pika cites (first among open models, seventh overall in the DesignArena text-to-video ranking) is Pika's own claim and a snapshot at launch.
A few things are still open. How well does a performance hold up when references change across a longer sequence? And what does the licence allow for commercial work? That's exactly what we want to know before a model like this goes beyond look tests for us. Pika first gave Flex to developers through its API programme and announced it for its own product from 2 November.