How to Run Pika H3 Flex Locally: ComfyUI, Diffusers and Hardware Requirements

Pika released Pika H3 Flex with open weights, so you can run it on your own hardware instead of a hosted service. This guide summarizes the official ComfyUI workflow and the Hugging Face release so you know what to download, where it goes and what hardware to expect before you start a 200+ GB download.

Everything here comes from the ComfyUI tutorial and the Pika-Labs/Pika-H3-Flex model card. File names and versions change quickly after a launch — always double-check them against those pages.

First: what the open release actually is

Pika H3 Flex is not a standalone checkpoint. The release is a set of LoRA adapters that sit on top of the MiniMax H3 base model (33B parameters, pinned to a specific base revision in the model card). To run Flex you need:

• the MiniMax H3 base components (transformer, text encoder, video VAE, audio VAE), and
• the Flex adapters for the task you want (first/last-frame or reference-to-video).

Sizes reported for the release: roughly 210 GB of base components plus about 5.5 GB of adapters; the Hugging Face repo lists about 220 GB in total.

Hardware requirements

Pika does not publish a hard minimum, but the signals are clear:

• The model card says representative 480p and 768p CLI requests passed on an H100 80GB.
• Larger or longer requests may require --cpu-offload.
• ComfyUI’s guide says a 5-second clip at about 768p finishes in roughly a minute on a single card (it does not name the GPU).

In practice, plan for an 80 GB datacenter GPU, fast NVMe storage with 250+ GB free, and plenty of system RAM if you offload. Consumer 24 GB cards are not a realistic target for full-quality runs. Native output is 480p or 768p — the 1080p/2K/4K upscaler is only on Pika’s hosted service.

Option 1: ComfyUI

Version

ComfyUI ships three Pika H3 Flex templates under Template Library → Video:

Model files and folders

The base files are shared with ComfyUI’s existing MiniMax H3 templates; only the two Flex adapters are new.

The adapter file names differ slightly between the ComfyUI docs and the comfy/ folder of Pika’s Hugging Face repo, so use whichever name your template’s loader node expects.

Sampler settings

The six-step schedule is what makes Flex fast, and it is easy to break:

• Load the adapter with the standard LoraLoaderModelOnly node at strength 1.0.
• Sampler lcm, sigmas from ManualSigmas, CFG 1.0 through BasicGuider.
• Six steps — keep the template’s sigma list (it needs at least seven values).
• Don’t stack turbo or “Lightning” LoRAs on top.
• Pika notes that generic six-step Euler sampling is not equivalent to its schedule.

Inputs and limits

• Reference-to-video accepts up to 9 images, 3 videos and 3 audio clips (12 files). Each video or audio clip must be 2–15 seconds, and each type’s combined length must stay at or under 15 seconds. At least one image or video is required.
• Durations are 5–15 seconds at 24 fps.
• Audio (dialogue, SFX, music) is generated in the same pass as the picture.

Option 2: Diffusers / Python

The model card shows a diffusers path that loads the base model and then the Flex LoRA:

import torch
from diffusers import DiffusionPipeline
from diffusers.utils import export_to_video

pipe = DiffusionPipeline.from_pretrained(
    "MiniMaxAI/MiniMax-H3", dtype=torch.bfloat16, device_map="cuda"
)
pipe.load_lora_weights("Pika-Labs/Pika-H3-Flex")
# see the repo's examples/README.md for the six-step schedule and task options

Install with pip install -U diffusers transformers accelerate. The snippet on the card is minimal; the repo’s examples/ folder documents the full sampling settings and the task names (t2va, i2va, l2va, fl2va, ref2va).

License check before you deploy

The adapters are model derivatives of MiniMax H3 and are released under the MiniMax H3 Community License Agreement; the example code is Apache 2.0. Community license terms can restrict usage, commercial scale or deployment regions — read the full license in the repo’s licenses/ folder before using local generations commercially.

Local vs hosted: which should you choose?

If you just want clips, the hosted route is faster and usually cheaper: open the generator or read the pricing breakdown.