Pika H3 Flex

Pika H3 Flex is an open-weight AI video model from Pika that makes clips with synced audio from text, frames or references, for creators and developers.

What it is: Pika's tuned version of MiniMax H3

Pika H3 Flex was launched on October 29, 2026. It is not a new base model. Pika post-trained MiniMax H3 and released the result as two small adapters that sit on top of the MiniMax H3 base. One adapter handles text and first or last frames, the other handles references. Pika says it cut the generator down to six sampling steps, which is why it can run faster than real time on its own hosted service.

• Text to video with synchronized stereo audio, 32 kHz, at 24 fps.
• First-frame, last-frame, or first-and-last-frame conditioning.
• Reference-guided generation with up to 12 files: 9 images, 3 videos and 3 audio clips, at least one image or video. Video and audio references must each be 2 to 15 seconds.
• Eleven aspect ratios from 21:9 to 9:21, and prompts up to 8,000 characters. You can describe hard cuts, camera moves and dialogue inside one prompt.
• Output is MP4 with H.264 video and AAC audio.

What you can run yourself, and what only Pika hosts

• Local open-weight inference makes native 480p or 768p video with mixed stereo audio.
• 1080p, 2K and 4K output, audio stems and prompt enhancement are Pika-hosted API features. The weights and GitHub code do not include them.
• Pika's page says the request generates at 768p and upscales once inside the same job to reach 1080p, 2K or 4K.
• The model is also listed for ComfyUI and Diffusers. ComfyUI's guide says a 5-second 1344x768 clip finishes in about a minute on a single card.

If you want to run it on your own GPU

• You need Python 3.12 and an NVIDIA GPU. The repo's tested cases ran on one H100 80GB per request. H200 and B300 are untested, and the package does not split one request across several GPUs.
• A default 480p, 5 second request used about 66 to 69 GiB of GPU memory in Pika's tests. A default 768p, 15 second request did not fit on an H100, and needed --cpu-offload, which drops GPU use to roughly 29 to 32 GiB but takes about 94 GiB of host memory.
• Allow at least 211 GB of disk for the MiniMax H3 base components plus about 5.5 GB for the two adapters.
• The command and Python package are called pika-h3-pro, even though the repository and model are named Pika H3 Flex. Search for that name if the installer output looks unfamiliar.
• The README says local speed will not necessarily match the hosted API, and that its six showcase videos are creative examples, not release-validation renders or throughput benchmarks.

License: check where you live before you download

• The adapters are Model Derivatives of MiniMax H3 and fall under the MiniMax H3 Community License, not an MIT or Apache license. GitHub shows the license as "Other".
• The license covers the whole world except the European Union, the United Kingdom, the Republic of Korea and the United States. Using the model or its outputs outside the covered area is not authorized.
• If your commercial products and services earn more than 20 million US dollars a year, you need written authorization from MiniMax first.
• If you offer others a product or service that generates outputs with it, you must bind users to terms at least as protective as the license's use restrictions and keep reasonable safeguards against misuse.
• Distributing copies to third parties needs the MiniMax NOTICE file. Pika asks, without requiring it, for a credit and citation.

Pricing and numbers, as Pika reports them

• Pika's technical blog lists a starting cost of $0.01 per output second for text and image to video, and $0.03 per second for reference to video. Its comparison table lists MiniMax H3 at $0.08 per output second. Seailife did not check the live pricing page.
• Pika says Flex ranks first among open-source text-to-video models, beats competitors by up to 91% in its own human evaluation, and runs 1.8x faster than H3 Max. These are the company's claims, not independent results.
• The product page and the blog disagree in places. The page says 3 to 15 seconds and three audio stems, dialogue, music and effects. The repo says 5 to 15 second requests, and the blog says five stems, adding foley and ambience.

Who it suits

Pika's own examples are multi-shot stories, motion graphics, product films and reference-guided animation, such as a frog table-tennis rally with five scripted camera moves and a four-shot sitcom directed from one image. The prompts and input assets for six showcases are published in the repo's docs. Seailife has not run the model.