Pika H3 Flex VRAM Requirements: How Much GPU Memory You Need to Run It Locally
PikaH3Flex EditorialPika H3 Flex needs about 66–69 GiB of GPU memory for a 5-second 480p clip and ~211 GB of disk. Tested GPUs, CPU offload numbers, and whether a 24 GB card can run it.
Running Pika H3 Flex locally takes an 80 GB-class GPU. In the release’s own tests, a default 5-second 480p clip allocated about 66–69 GiB of GPU memory during sampling, and the representative cases were run on an NVIDIA H100 80GB. A 15-second 768p clip did not fit on the H100 without CPU offload. You also need at least 211 GB of disk for the base model plus about 5.5 GB for the Flex adapters. A 24 GB consumer card such as an RTX 4090 is far below these numbers.
All figures below come from docs/hardware.md in the official Pika-Labs/Pika-H3-Flex GitHub repository. They are measurements of representative runs, not guaranteed minimums.
GPU memory by task
| Request | GPU memory (sampling) | Fits on H100 80GB? |
|---|---|---|
| 480p, 5 s, default mode | ~66–69 GiB | Yes |
| 768p, 15 s, default mode | More than 80 GB | No |
768p, 15 s, with --cpu-offload | ~29–32 GiB GPU, ~94 GiB host RAM | Yes |
Two caveats from the repo: the text encoder and VAEs have separate memory needs that the sampling figure excludes, and you should leave headroom for references, long prompts and output buffers.
What was actually tested
The repo lists the cases that passed on an H100 80GB:
| Task | Resolution | Duration | Memory mode |
|---|---|---|---|
| Text to video | 480p | 5 / 15 s | Default |
| First, last or both endpoint frames | 480p | 5 s | Default |
| Text, image or both endpoint frames | 768p | 5 s | Default |
| Image references | 480p / 768p | 5 s | Default |
| Image + video + audio references | 480p | 5 s | Default |
| Text or image references | 768p | 15 s | CPU offload |
H200 and B300 were not qualified at release. The package uses one GPU per request: extra GPUs can run separate jobs, but one clip is not split across several cards.
Disk, RAM and software
- Disk: at least 211 GB for the MiniMax H3 base components, plus ~5.5 GB for the two Flex adapters, plus room for dependencies and download staging. Fast NVMe helps.
- System RAM: with CPU offload, the 768p/15 s run used about 94 GiB of host memory including page cache.
- Software: Python 3.12, a supported NVIDIA driver and the pinned package versions. Pip, Conda and Docker installs were exercised.
Can I run Pika H3 Flex on an RTX 4090, 3090 or 5090?
Not with the official settings. The smallest tested job used roughly 66 GiB of GPU memory, more than double what a 24 GB or 32 GB card has, and CPU offload still needed about 29–32 GiB of GPU memory for the 768p run. Community quantized or heavily offloaded setups may appear, but they are not part of the official release and will trade speed and quality. For consumer hardware, a cloud option is the practical route.
Can I run it on Google Colab or a rented GPU?
Free Colab GPUs are far too small. A rented 80 GB cloud GPU works, but you pay for the hours spent downloading ~216 GB, setting up and idling, not just for generation. Local output also stops at 768p: 1080p, 2K and 4K upscaling is only available on hosted services.
No GPU? Use Pika H3 Flex online
If you don’t have an 80 GB card, you can use Pika H3 Flex in the browser instead: no install, no download, 480p to 4K, with audio. Try text to video or image to video; a 5-second 768p clip is 20 credits (see pricing). If you have the hardware, follow the ComfyUI and Diffusers setup guide or the GitHub and Hugging Face guide.
FAQ
How much VRAM does Pika H3 Flex need? About 66–69 GiB of GPU memory for a default 5-second 480p clip, measured on an H100 80GB. Longer or 768p clips need more, or CPU offload.
What is the minimum GPU for Pika H3 Flex? The release was tested on an NVIDIA H100 80GB. Plan for an 80 GB-class datacenter GPU.
Can Pika H3 Flex run on 24 GB of VRAM? Not with the official release. Even the CPU-offload mode used about 29–32 GiB of GPU memory for a 768p clip.
How much disk space does Pika H3 Flex need? At least 211 GB for the base model plus about 5.5 GB for the adapters, and extra space for dependencies.
Can I use multiple GPUs for one clip? No. The package runs one request per GPU; more GPUs let you run more clips in parallel.
How do I use Pika H3 Flex without a GPU? Use it online in a browser studio such as this site. Everything runs in the cloud.