# Pika H3 Flex makes video in six steps instead of 50, and the adapters can be downloaded

URL: https://ai-pulse.aglaganova.com/blog/pika-h3-flex-makes-video-in-six-steps-instead-of-50-and-the-adapters-can-be-downloaded
Published: 2026-11-04
Author: Roman Aglaganov, Full-stack developer
Category: Models
Region: North America
Tags: open-weights

> Built on MiniMax’s open H3 model, Flex generates clips with synchronised stereo sound, takes up to 12 reference files and can edit existing footage. Running it yourself still needs a data-centre GPU.

![Illustration: a strip of grey film frames receding to the right, with six black frames in the foreground and the last one marked red](https://ai-pulse.aglaganova.com/api/media/file/pika-h3-flex-1.png)

Pika, the AI video start-up, released H3 Flex on 29 October, a version of MiniMax’s open video model H3 that has been retrained to generate in six steps rather than 50. “Your first generation is a starting point. The creative work is what happens next,” the company writes, and the release is aimed at people who make many variations of a shot.

## What it does

**Speed:** Pika says Flex can produce a 15-second, 768p clip with synchronised audio in as little as seven seconds of model compute. The company is careful to say that this is a controlled benchmark, not how long a user will wait, since references, upscaling and service overhead all add time.

**References:** A single generation can take up to 12 images, videos or audio files, each given a role such as character, setting or motion. Flex can also edit existing footage, changing elements like clothes, props or the setting while keeping the performance and timing.

**Sound:** Clips come with synchronised stereo audio, and through Pika’s API users can get five separate tracks: dialogue, score, sound effects, Foley and ambience.

Pika says Flex ranked first among open-weights models and seventh overall on the DesignArena text-to-video leaderboard at launch, a snapshot that will change.

## How open it is

Pika has published the generator adapters and inference code on Hugging Face and GitHub, for text-to-video, first-and-last-frame and reference-to-video generation at 480p and 768p. They are licensed under MiniMax’s H3 Community License rather than a standard open-source licence, and the base model components take up about 210 GB. Pika says it tested typical requests on a single Nvidia H100 with 80 GB of memory. Upscaling to 1080p, 2K or 4K, the separate audio tracks and Pika’s fast serving setup remain available only through its own service. Flex is available through the Pika API now and was due to reach Pika’s app on 2 November.

## Sources

- [Pika: Pika H3 Flex](https://pika.art/blog/pika-h3-flex)
- [Hugging Face: Pika-Labs/Pika-H3-Flex model card](https://huggingface.co/Pika-Labs/Pika-H3-Flex)