Pika H3 Flex GitHub and Hugging Face: Download, Install and Run the Open Weights

Pika H3 Flex has two official homes: the inference code on GitHub at Pika-Labs/Pika-H3-Flex and the weights on Hugging Face at Pika-Labs/Pika-H3-Flex. The GitHub repo gives you the pika-h3-pro command-line tool; the Hugging Face repo holds the Flex adapters plus a pinned copy of the MiniMax H3 base files they run on. This guide explains what is in each, how to install it, and what you need before you start a 216 GB download.

Commands and numbers below are taken from the repo README, docs/hardware.md and the model card as of November 1, 2026. The project is days old, so check the repos for changes before you copy anything.

The two repos at a glance

What is in the Hugging Face repo

Pika H3 Flex is not a standalone model. It is a pair of LoRA adapters for the 33B-parameter MiniMax H3:

• fl2va/ — FP32 LoRA (rank/alpha 128) for the base transformer. Powers text-to-video and all start/end-frame tasks.
• ref2va/ — FP32 LoRA (rank/alpha 128) for the base transformer_ref. Powers multimodal reference-to-video.
• Base components in Diffusers format — transformer/, transformer_ref/, text_encoder/, vae/, audio_vae/, tokenizer and scheduler configs — about 210 GB.
• comfy/loras/ — BF16 copies of the adapters in ComfyUI format.

The adapters themselves are only about 5.5 GB; almost all of the download is the base model.

What is in the GitHub repo

The repo (src/pika_h3/, tools/, patches/, examples/prompts/, docs/, a Dockerfile and environment.yml) packages a single CLI called pika-h3-pro. It downloads the pinned base revision and the adapter a task needs, then writes an MP4 plus a JSON file with the prompt, seed, geometry and timings.

Requirements listed in the README:

• Python 3.12 and an NVIDIA GPU with a compatible driver.
• git, patch, ffmpeg and a C compiler with headers (Triton needs it at runtime).
• At least 211 GB of disk for base files, 5.5 GB for adapters, plus room for dependencies and download staging.
• An HF_TOKEN at runtime if Hugging Face asks for one.

Install: pip, Conda or Docker

pip

git clone https://github.com/Pika-Labs/Pika-H3-Flex.git
cd Pika-H3-Flex
python -m pip install -r requirements.txt
python tools/apply_diffusers_patches.py
python -m pip install --no-deps -e .
pika-h3-pro --help

Conda

conda env create -f environment.yml
conda activate pika-h3-pro
python tools/apply_diffusers_patches.py
python -m pip install --no-deps -e .

Docker

docker build -t pika-h3-pro .
docker run --rm --gpus 'device=0' pika-h3-pro --help

Do not bake your Hugging Face token into a Docker image layer; pass it at runtime.

Run your first clip

Pin the release so your results are reproducible:

export MODEL_REPO=Pika-Labs/Pika-H3-Flex
export MODEL_REVISION=0ddf55ad7067ca1e76812253bd7b4600df87f759

Text to video at 480p, 5 seconds:

pika-h3-pro --task t2va --prompt-file examples/prompts/t2va.txt \
  --model-repo "$MODEL_REPO" --revision "$MODEL_REVISION" \
  --resolution 480p --aspect 16:9 --duration 5 --seed 42 \
  --output outputs/sample.mp4

A 15-second 768p clip needs CPU offload on an 80 GB card:

pika-h3-pro --model-repo "$MODEL_REPO" --revision "$MODEL_REVISION" \
  --task t2va --prompt-file examples/prompts/t2va.txt --resolution 768p \
  --duration 15 --seed 42 --cpu-offload --output outputs/long-768p.mp4

Tasks and flags

Useful flags: --prompt-file (UTF-8, 1–8,000 characters), --resolution (480p or 768p), --aspect (11 ratios), --duration (5–15), --seed, --first-image, --last-image, --ref image:subject.png (repeatable; up to 9 images, 3 videos and 3 audio clips) and --cpu-offload.

Hardware: what actually fits

From the repo’s docs/hardware.md:

One request runs on one GPU; extra GPUs only help by running separate requests. H200 and B300 are listed as untested. Native output stops at 768p — the 1080p, 2K and 4K upsampler, audio stem separation and prompt enhancement are Pika-hosted and not in the repo.

Want to try it on a smaller card? ComfyUI ships quantized base files (int8 and NVFP4) that cut memory a lot; see our ComfyUI and local setup guide and the Pika H3 Flex LoRA guide.

Diffusers instead of the CLI

The model card also shows a Diffusers path — load the base and then the Flex LoRA:

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained(
    "MiniMaxAI/MiniMax-H3", dtype=torch.bfloat16, device_map="cuda"
)
pipe.load_lora_weights("Pika-Labs/Pika-H3-Flex")

The full six-step settings live in the repo’s examples/README.md, pinned to a specific upstream Diffusers commit in examples/requirements-diffusers.txt. Using a different sampler schedule breaks the speed-up — see how the 6-step sampler works.

License in one minute

• Code (GitHub): Apache 2.0.
• Adapters (Hugging Face): model derivatives of MiniMax H3 under the MiniMax H3 Community License Agreement, with a required NOTICE file.
• Qwen text encoder and tokenizer assets: Apache 2.0.

“Open weights” is not the same as unrestricted use. Read the license in the repo’s licenses/ folder before shipping local generations commercially.

Do you need the repos at all?

Use them if you want private, self-hosted generation and have an 80 GB-class GPU. If you only want clips, the hosted route is quicker and includes 1080p–4K: try Pika H3 Flex text to video or image to video in the browser, or compare every option in how to use Pika H3 Flex.

FAQ

Is Pika H3 Flex on GitHub? Yes. The inference code is at github.com/Pika-Labs/Pika-H3-Flex; the weights are on Hugging Face.

How big is the download? About 216 GB: roughly 211 GB of MiniMax H3 base files and 5.5 GB of Flex adapters.

What GPU do I need? Pika tested on a single H100 80GB. A 5-second 480p clip used about 66–69 GiB of GPU memory.

Can the open weights make 4K? No. Local output is 480p or 768p; 1080p to 4K come from Pika’s hosted upsampler.