Pika H3 Flex is a video model Pika released on October 29, 2026. Fine-tuned from MiniMax H3, it cuts sampling from 50 steps to 6 and generates 5- to 15-second clips with synchronized stereo audio in one go. Ordinary users can just go to Pika's website (pika.art); it has been in the main app since November 2. Anyone wanting to run it locally can download the weights from Hugging Face, but read the license first.
How ordinary users use it: free quota for two weeks after launch
Open Pika’s website (pika.art), pick H3 Flex, enter text, or supply first and last frames, reference images, reference videos, and reference audio — up to 12 items (9 images, 3 videos, 3 audio clips) — to generate a 5- to 15-second clip. The reference mode is its selling point: one product image plus a reference audio clip is enough to produce a short film with dialogue or music. For the first two weeks after launch there are 15 free generations a day and prices are halved, ending around November 12; claim it first, then try things out at leisure.
Running it locally: check the hardware and the license first
The weights are in the Hugging Face repository Pika-Labs/Pika-H3-Flex, and the inference code is on GitHub under Apache-2.0. Locally it is two LoRA adapters mounted on the MiniMax H3 base, whose files total around 210GB. Local output tops out natively at 768p; 4K finishing and five separate audio stems are Pika-hosted features. Running it takes an H100-class GPU, so do not force it on a consumer card. More important still: the weights fall under the MiniMax H3 Community License, whose standard grant excludes local deployment in the US, EU, UK, and South Korea. Before commercial use, read the license text itself, and if in doubt, use the hosted API.
What "7 seconds" really means
The 7 seconds Pika talks about is the model's own compute time on its servers, not the total time from clicking generate to receiving the finished clip. Queuing, uploading reference files, and post-processing all add more time. The 6-step sampling is real, and so is the speed advantage, but do not plan a workflow around 7 seconds. If you need 1080p or higher, or want to mix stems separately, go straight to the hosted service instead of tinkering locally.
