Pika H3 Flex: a fast open-weight video model with 12 references and split audio stems

Pika has released Pika H3 Flex, a video generation model it post-trained from MiniMax H3, and it's putting the model's weights out in the open. The pitch is speed and control: the company says Flex can generate a 15-second clip with synchronized audio at 768p in as little as seven seconds of model compute, because it now needs six generation steps instead of fifty.

What it can do

According to Pika's announcement:

Up to 12 reference files per generation, mixing images, video and audio, so a single shot can draw on things like a character design, a location and a sound reference at once.
First- and last-frame control, so you can pin where a shot starts and ends.
Five separate audio stems through the API: dialogue, score, sound effects, foley and ambience.
Drafts at 480p or 768p, with a hosted upscaler that takes a chosen clip to 1080p, 2K or 4K.

At launch, Pika says Flex was the top open-weight model and seventh overall on DesignArena's text-to-video leaderboard.

What is actually open

The open release covers text-to-video, first- and last-frame generation and multimodal reference-to-video at 480p and 768p, with the generator adapters and inference code on Hugging Face and GitHub. The upscaler to higher resolutions stays on Pika's own hosted service. Flex is available through the Pika API now and comes to Pika's main creation app on November 2.

My take

The part that stands out to me as someone who follows animation is the reference and frame control. Feeding in character designs and fixed start and end frames is close to how traditional pre-production works, and it's also why tools like this keep sparking arguments in the anime world about credit and consent for artists' work. Open weights mean more people can run and study the model themselves, which cuts both ways. Either way, sound split into separate stems is a genuinely practical touch for anyone editing short videos.

Read Pika's announcement for the demo clips.