Pika-Labs/Pika-H3-Flex

Pika H3 Flex is an inference package and released adapter pair for a video-and-audio generation model. It suits creators or engineers who want to run the model themselves on their own GPU rather than call a hosted service, and who accept a large download and H100-class hardware. The adapters are model derivatives of MiniMax-H3, so the model licence, not just the code licence, governs their use.

What it produces

One model family with two task adapters: FL2VA handles text-only and endpoint-frame conditioning (t2va, i2va, l2va, fl2va), while Ref2VA takes ordered multimodal references. Inputs can include text, first/last images, up to 9 images, 3 videos and 3 standalone audio clips, capped at 12 reference files, with each video or audio source 2–15 seconds. Native output is 480p or 768p across 11 aspect ratios, 24 fps, 5–15 requested seconds (rounded to a 17n+5 frame grid, so 5 s yields 124 frames), with 32 kHz stereo audio mixed into an H.264/AAC MP4 plus JSON metadata. Higher resolutions (1080p, 2K, 4K), audio stems and prompt rewriting are separate hosted Pika API endpoints, not part of local inference (README).

How it works

The pipeline encodes the prompt and media, generates video and audio latents with the selected adapter, then decodes and combines them into the MP4. Generation uses six denoising forwards with a specific knot schedule, fresh intermediate noise, video shift 12 and audio shift 3; the README warns that generic six-step Euler sampling is not equivalent, and results can still differ across attention kernels, devices and library versions (README). The package downloads pinned MiniMax-H3 components and adapters automatically, and validates adapter hashes and sampling parameters before loading. Metadata JSON records the final prompt, seed, geometry, offload setting and timings, so a prompt you consider private ends up in that file.

Requirements and constraints

Local use targets Python 3.12 on an NVIDIA GPU with a compatible driver; pip installation needs system git, patch, ffmpeg and a C compiler with development headers, and Conda and Docker recipes supply these. Representative tests use one H100 80GB per request; H200 and B300 are untested. One request runs on one GPU — exposing eight GPUs does not shard it, though separate processes can use idle cards. Both tasks together downloaded about 210 GB of base files plus 5.5 GB of adapters, before dependencies and staging. --cpu-offload keeps transformer weights on the CPU to lower sampling VRAM at the cost of host RAM and transfer time, but does not offload the text encoder or VAEs, so it is not a minimum GPU requirement. The tested verification table does not qualify every resolution, duration or reference combination, and maximum-size reference sets may exceed memory (hardware docs).

Licensing

Inference and loader code is Apache-2.0, but the adapters are distributed under the MiniMax H3 Community License Agreement with a required notice (repository). Anyone deploying the model must check that licence, not just the code licence.