Pika Open-Sources H3 Flex: 15-Second Videos With Music in ~7 Seconds

Much of video generation's progress over the past two years has been bought with waiting: better quality, but minutes per clip. Pika is going the other way and making speed the headline.

Pika has released H3 Flex, a video model post-trained on MiniMax H3, with open weights. By the company's account it can produce a 15-second, 768p clip with synchronized audio and music in about seven seconds of transformer compute. That is shorter than the clip itself, hence "faster than real time."

From 50 steps to six

The speed comes from distillation. Pika fine-tuned H3 on an internally curated, aesthetics-focused dataset, used distribution matching distillation to cut denoising from 50 steps to six, then used reward models and preference optimization to restore quality. Pika says Flex is 1.8x faster than H3 Max and ranks first among open-weight models and seventh overall on DesignArena's text-to-video leaderboard.

Speed matters for more than saving time. Creators never get a clip in one shot; they iterate on the shot, the line, the pacing. When each generation drops from minutes to seconds, the workflow changes: it feels like scrubbing a timeline rather than placing an order and waiting.

Control for professionals

Beyond speed, Flex is pitched on control. It accepts up to 12 references (nine images, three videos, three audio clips), so you can transfer motion from a reference video or set a voice from a sample. A single prompt can direct hard cuts, camera moves and dialogue across multiple shots.

Audio is the other standout. Flex can export the soundtrack as five separate stems, dialogue, score, SFX, foley and ambience, which Pika calls a first for a video model. For anyone heading into post-production that beats a single stereo mix: don't like the music, swap it out without regenerating the whole clip.

Output runs from 480p to 4K; higher resolutions are generated at 768p and upscaled once in the same job. API pricing ranges from $0.01 per second at 480p to $0.08 at 4K, cheap for this class.

Open, with a few things held back

The scope of the open release is worth spelling out. Pika published the generator adapters (two LoRAs on the H3 base) and inference code, covering text-to-video, first/last-frame and multimodal reference-to-video with synchronized audio at 480p and 768p. The 1080p-to-4K upsampler, stem separation, prompt enhancement and accelerated serving stack stay on Pika's hosted service. The license follows the MiniMax H3 Community License, so check its territorial and commercial terms before using outputs commercially.

It is an increasingly common playbook: open the model to win community and ecosystem, keep the last mile that defines the experience in-house. Researchers and ComfyUI tinkerers get plenty; commercial users chasing finished quality will still find Pika's API the easier path.

Pika's positioning shift is telling too. A company that once broke out with its own models and consumer app is now building on MiniMax's base and competing on post-training and product. As base video models get ever more expensive to train from scratch, base model plus post-training plus workflow may become how mid-sized players survive.