Pika releases open-weight video model with six-step generation and synchronized audio

Pika released H3 Flex on October 29, 2026, giving developers downloadable adapters and code for generating video with synchronized sound. Built through further training of MiniMax H3, the model targets a familiar creative task: revising a promising first attempt. Its core generation runs locally, while higher-resolution finishing and separate audio tracks remain tied to Pika’s service.

Pika started with MiniMax H3 rather than building a new model from scratch. It says it reduced generation from 50 steps to six, then refined the accelerated model for quality and creative control. The downloadable adapters are additional trained weights: the Hugging Face loading example starts with MiniMax H3 and then loads Flex’s weights.

Pika’s headline speed result is a 15-second clip with synchronized audio at 768p, requiring as little as seven seconds of transformer computation. That is a controlled benchmark, not the time from submitting an API request to receiving a finished file. Reference processing, higher-resolution finishing, audio separation, hardware and service overhead can all affect completion time.

For quality, Pika says Flex placed first among open-weight entries and seventh overall on DesignArena’s text-to-video leaderboard in its launch snapshot. The ranking describes that moment, not a permanent position.

Flex accepts up to 12 reference files, mixing images, video and audio in one generation. Pika describes using one image to establish a character, another for the setting, and a video for movement. Prompts specify each reference’s contribution. The released modes also cover text prompts and control through a clip’s first or last frame.

• Composition: combine drawn keyframes with photographic references to guide framing, camera choices and visual style.
• Movement: use reference choreography while generating a different character and setting.
• Selective editing: request changes to clothing, props or surroundings while asking Flex to preserve performance and timing.

Pika’s finishing workflow starts with a 768p draft, then upscales the chosen version to 1080p, 2K or 4K. Through the API, users can request five separate audio tracks: dialogue, score, sound effects, Foley and ambience. These let editors lower the music without lowering speech, or adjust effects independently.

The model page lists the MiniMax H3 Community License Agreement. Pika says developers and researchers can run, customize and build tools around the release under its license terms; local performance depends on hardware and setup.

Hosted access is already available through Pika API Club. The next scheduled step is November 2, 2026, when Pika says Flex will reach its creation product. That consumer rollout remains a plan, rather than current availability.