Pika Releases Pika H3 Flex, an Open-Weights Video-with-Audio Model Built on MiniMax H3
Pika has released downloadable adapters for Pika H3 Flex, a model that generates video and synchronized audio in six steps, post-trained on MiniMax H3. The local model outputs in 480p or 768p; 4K and separate audio stems require Pika's paid API.
Pika released Pika H3 Flex this Thursday, October 29, at 6:37 p.m. Paris time, a video-with-synchronized-audio generation model whose weights are available to download from day one on Hugging Face, with code on GitHub and support in diffusers and ComfyUI. Pika positions it as a rapid iteration tool for professional creators.
A Model Derived from MiniMax H3
Pika H3 Flex is not a model trained from scratch. According to the GitHub repository, it is a post-training of MiniMax H3, distributed as two adapters: one for text and start or end frame images, and the other for multimodal references. It generates video and stereo audio in six sampling steps.
Locally, the model outputs 480p or 768p, across 11 aspect ratios, for clips lasting 5 to 15 seconds at 24 frames per second. It accepts up to 12 references: 9 images, 3 videos, and 3 audio clips.
Two factors limit its use on consumer hardware:
- the license: the adapters fall under the MiniMax H3 Community License Agreement, rather than a standard open-source license; only the inference code is under Apache 2.0;
- the hardware: one NVIDIA GPU per request, test cases run on an 80 GB H100, and roughly 210 GB of base files to download, plus 5.5 GB of adapters.
What Remains Behind the Paywall
The features most useful for production require Pika’s hosted API: upscaling to 1080p, 2K, and 4K from a draft, exporting up to five separate audio stems (dialogue, music, sound effects, foley, ambient), and prompt enhancement. The API is billed per second of generated video, starting at $0.01 per second in 480p at introductory rates ($0.02 at standard pricing). Pika API Club members receive 15 free generations per day through November 5.
Claims to Verify
Pika claims its model can generate a 15-second clip in seven seconds and is faster than H3 Max, noting that these measurements measure model compute time rather than total turnaround time. The developer also claims it ranks 7th in video quality on DesignArena’s text-to-video leaderboard—ahead of MiniMax H3 and H3 Max—and first among open-weights models. We have not verified this on DesignArena.
Our take
Pika H3 Flex illustrates a strategy becoming common in generative video: release the open weights of a “draft” base model to win over the community (ComfyUI, diffusers), while keeping high-end finishing (4K, separate audio stems) inside a paid API. For studios and developers, this is meaningful progress: a multi-reference video-with-sound model that can be run and fine-tuned locally. However, it remains bound to MiniMax’s license, whose terms warrant scrutiny before any commercial deployment, and it demands a data-center GPU. This is not a model built for a laptop.
Sources
- Pika (X) — Pika H3 Flex announcement Primary source
- Pika (X) — code, weights, diffusers, and ComfyUI Primary source
- Pika — Pika H3 Flex page (features and pricing) Primary source
- GitHub — Pika-Labs/Pika-H3-Flex Primary source
- Hugging Face — Pika-Labs/Pika-H3-Flex Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
