Pika drops H3 Flex: video and audio in just six steps

Pika H3 Flex is an open-weight video and audio generation model post-trained on top of MiniMax-H3. Its standout feature is raw iteration speed: it generates video alongside synchronized 32 kHz stereo audio in just six sampling steps, eliminating the need to render frames first and stitch sound effects later.

The model natively supports 480p and 768p outputs across 11 aspect ratios, ranging from cinematic 21:9 and widescreen 16:9 to vertical 9:16 and ultra-tall 9:21. Output durations range from 5 to 15 seconds at 24 fps. Because frame counts align to the underlying grid, a 5-second request yields 124 frames (about 5.17 seconds), while a 15-second clip renders 362 frames.

Two adapters for flexible control

Instead of generic Euler sampling, Pika H3 Flex relies on a tuned six-step denoising schedule paired with two specialized task adapters: FL2VA and Ref2VA.

FL2VA manages text prompts and endpoint conditioning. It lets you run pure text-to-video-and-audio, guide motion from an opening frame, anchor the scene to an ending frame, or lock both boundaries to interpolate realistic motion and acoustics between them.

Ref2VA provides deep multimodal reference control. It accepts up to 12 reference files in a single pass, combining up to 9 images, 3 video clips, and 3 standalone audio tracks. Individual clips must be between 2 and 15 seconds, and embedded video audio is preserved automatically, giving creators fine-grained leverage over character motion and scene atmosphere.

Hardware specs and setup

Local inference demands enterprise hardware. Pulling the required weights requires roughly 210 GB for base components and 5.5 GB for the adapters, plus scratch space for staging.

Pika tested representative workloads on a single NVIDIA H100 80GB GPU. The current CLI distribution runs one task per GPU and does not shard across multiple cards, though multiple idle GPUs can handle independent generation processes in parallel.

To prevent out-of-memory crashes on heavy prompts, the tool includes a CPU offload mode. This keeps Transformer weights in system RAM and streams blocks to VRAM on demand. The trade-off is extra transfer latency, but it enables 768p, 15-second generation runs on an 80GB accelerator.

The codebase is available on GitHub under Pika-Labs/Pika-H3-Flex with an Apache-2.0 license, while the adapters follow the MiniMax H3 Community License. Setup uses the pika-h3-pro command line via Python 3.12, Conda, or Docker. Higher resolutions such as 1080p, 2K, and 4K, along with separated audio stems, remain hosted features accessed through Pika's commercial API.

Why it matters

For commercial directors, concept animators, and visual effects artists, Pika H3 Flex turns video generation into a genuine rapid-prototyping loop. Slashing inference down to six steps means teams can block out scene choreography, pacing, and basic sound design without waiting minutes per test render.

By baking native stereo generation directly into the diffusion pipeline, it bypasses post-generation audio synchronization pipelines. The catch is hardware accessibility: requiring an 80GB enterprise GPU keeps it out of reach for hobbyists on local consumer hardware, making it strictly a power tool for well-equipped studios and AI engineers.