Pika releases H3 Flex, an open-weight video model that generates synchronized audio in six sampling steps

Pika Labs has released Pika H3 Flex, a post-trained version of its H3 model that generates video with synchronized audio in six sampling steps. The model is listed on Hugging Face with downloadable weights, and Pika describes it as built for professional iteration.

Key specifications

• Base: post-trained H3 model
• Sampling: six steps, with synchronized audio
• Reference inputs: up to 12, with shot direction through references and prompts
• Local open-weight inference: native 480p and 768p, with mixed stereo audio
• Pika API: 1080p, 2K and 4K video generation, audio stems and prompt enhancement
• Parameter count, context window, training cutoff and pricing: not yet disclosed

The model page also lists an omni-reference capability, along with links to model weights, video samples, a quickstart, a prompt guide and Pika API extensions.

Intended use

According to Pika, H3 Flex is designed to give professional creators the flexibility and speed to explore with precision. The company lists multi-shot storytelling, motion graphics, product films and reference-guided animation as the ideal use cases.

Sample prompts

The model page includes sample prompts with expandable generation descriptions. They show the control style the model is built around:

• Table-tennis rally: a prompt references an input image for the two players, outfits, frog mask, table and drained pool. It specifies five timed camera moves in a 15-second clip, including an overhead shot, a crane move, whip pans, a 180-degree orbit and a low-angle dolly. It also dictates natural real-time motion and a sound design that includes a drum-and-bass track.
• Apartment tower: a single continuous shot with no cuts. The prompt lays out second-by-second camera orbits, push-ins and booms around three people doing slow stretches on stacked balconies. It specifies diegetic sound only, with no music or speech.
• 1990s sitcom: a five-shot, 15-second scene with hard cuts, a live studio audience and two lines of dialogue. The prompt matches an input image.
• Motion graphic: a full-bleed 16:9 Swiss/geometric design built from a reference image. It is beat-synced to a 120 BPM track and spells out the word MOVE.

At least one further sample prompt on the page begins as a 15-second fast-cut cinematic sequence.

What this means

H3 Flex splits its capabilities between two tiers. The open weights cover native 480p and 768p with mixed stereo audio on local hardware. Resolutions of 1080p, 2K and 4K, plus audio stems and prompt enhancement, are available only through Pika's hosted API.

The six-step sampling count and the 12-reference limit show where the model is aimed: fast iteration and controlled output rather than single-prompt generation. The sample prompts support this. They combine reference images, timestamped shot lists, explicit camera moves and written sound direction in one request.

Pricing, parameter count and license terms are not stated in the material on the page, so cost and deployment requirements cannot yet be compared with other video models.