How to run FLUX 3 Klein locally

FLUX 3 Klein is an open-weight image model from Black Forest Labs (BFL). According to official BFL documentation, the model “fits on consumer GPUs, down to an RTX 3090” and is released in BF16, FP8, and NVFP4 weight variants. It uses the Qwen3-VL-4B-Instruct text encoder (licensed under Apache-2.0) and is available as a 50-step base checkpoint (with guidance set to 3.5, intended for fine-tuning and LoRA training) and a 16-step distilled checkpoint.

Before you can pull the model files to your machine, note that the weights are gated on Hugging Face. You must log in to your Hugging Face account and accept the FLUX Kommunity License v.1.0 before downloading. The ComfyUI files sit in a separately gated repository, black-forest-labs/flux-3-image-klein-comfy, where you must accept the license agreement as well. For full details on where the files are hosted, see our download guide.

Below are the ways to run FLUX 3 Klein on your own hardware.

Path	What you need	Official source
ComfyUI	An updated ComfyUI installation, downloaded weights, and the official workflow file.	BFL Documentation
Diffusers	Python environment with Hugging Face diffusers (the model card’s ModularPipeline setup installs it from Git), plus transformers for the prompt upsampler (or captions from another LLM).	Hugging Face Model Card
BFL Reference CLI	A local clone of the BFL repository installed in editable mode (pip install -e .).	Hugging Face Model Card
Mac	No official Apple Silicon statement found in BFL documentation; see our Mac page.	—

VRAM Before You Start

Before choosing a local setup path, verify that your graphics card has enough memory for the weight variant you want to run. The Hugging Face model card provides measured peak VRAM figures recorded on an NVIDIA B200 GPU generating at 1024x1024 resolution with the prompt upsampler (pup) turned off:

Weight Variant	Resolution	Prompt Upsampler (pup)	Measured Peak VRAM (B200)
BF16	1024x1024	Off	31 GB
FP8	1024x1024	Off	26 GB
NVFP4	1024x1024	Off	25 GB

Important note on the prompt upsampler (pup): FLUX 3 Klein comes with a 9B parameter prompt upsampler (pup), which is an 18.8 GB download. pup is on by default in the reference CLI and adds about 28 GB of GPU memory on top of the model weights (on an 80 GB GPU, per the model card); Diffusers runs no upsampler on its own. On a consumer GPU you will likely need to turn pup off, but the model card says: “Do not use the model without a prompt upsampler.” The model takes full scene captions, and a short prompt sent as typed gives generic output. If you turn pup off, supply a full scene caption written by another LLM using BFL’s system_prompts.json (--no-pup sends your prompt as typed).

If your GPU has less than 25 GB of VRAM, the ComfyUI model card notes that int8 ComfyUI files can run on 8 GB GPUs such as the RTX 3070. For a complete breakdown of memory usage across setups, read our VRAM and hardware requirements page.

Run with Diffusers

You can run FLUX 3 Klein in Python using the Hugging Face diffusers library. The model card’s documented ModularPipeline setup installs diffusers from Git with this command:

pip install git+https://github.com/huggingface/diffusers.git

Once installed, and after you have logged in to Hugging Face and accepted the gated repository license, you can load the model using ModularPipeline.from_pretrained:

import torch
from diffusers import ModularPipeline

# Load the FLUX 3 Klein pipeline
pipe = ModularPipeline.from_pretrained("black-forest-labs/flux-3-image-klein")
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

# Placeholder full scene caption. In practice, generate it with pup or
# another LLM using BFL's system_prompts.json, as in the model card's
# upsample() example; Diffusers runs no prompt upsampler on its own. pup is
# loaded via Transformers (Qwen3_5ForConditionalGeneration). See the model
# card for the full upsampler code and generation parameters:
# https://huggingface.co/black-forest-labs/flux-3-image-klein
caption = (
    "A small red fox sits on a moss-covered log in a misty pine forest at dawn. "
    "Soft golden light filters through the trees from the left, catching the fox's "
    "fur and the drifting fog. The fox looks toward the camera with its tail curled "
    "around its paws. Shot at eye level with a shallow depth of field, the background "
    "trees fade into a pale blue-green haze."
)
image = pipe(prompt=caption, output="images")[0]

Diffusers runs no prompt upsampler on its own, and the model card says: “Do not use the model without a prompt upsampler.” Either load pup via Transformers (Qwen3_5ForConditionalGeneration), which is an 18.8 GB download and adds about 28 GB of GPU memory (on an 80 GB GPU, per the model card), or write the full scene caption with another LLM using BFL’s system_prompts.json.

Run with ComfyUI

If you prefer a node-based graphical interface or need to use lower-VRAM int8 files, you can run FLUX 3 Klein in ComfyUI. The official BFL documentation outlines three steps:

1. Update ComfyUI: Make sure your local ComfyUI installation is updated to the latest version so it supports the FLUX 3 Klein architecture.
2. Download the weights: The ComfyUI files live in the separately gated repository black-forest-labs/flux-3-image-klein-comfy. Log in to Hugging Face, open that repository, accept the license agreement there, and then download the files. The ComfyUI templates use int8 single-file checkpoints, which run on 8 GB GPUs such as the RTX 3070; bf16 and nvfp4 versions are in the same repository.
3. Load the workflow: Download and load the official workflow from the ComfyUI Klein page.

For step-by-step file placement instructions, read our ComfyUI guide.

Run with the Reference CLI

Black Forest Labs provides a reference command-line interface (CLI) for running the model directly. To set it up, clone the official BFL repository and install the package in editable mode:

pip install -e .

Next, download the model weights into weights/ (the model card uses snapshot_download of the black-forest-labs/flux-3-image-klein repository), then fetch the safety classifier, which is always loaded, and the prompt upsampler:

hf download Falconsai/nsfw_image_detection --local-dir weights/nsfw-classifier
hf download black-forest-labs/flux-3-image-klein-pup --local-dir weights/pup

Then generate an image:

python scripts/cli.py '<prompt>' --model distilled --out out.png

The 9B prompt upsampler (pup) is on by default in the reference CLI. It is an 18.8 GB download and adds about 28 GB of GPU memory (on an 80 GB GPU, per the model card), so you can pass --no-pup to turn it off. The model card says “Do not use the model without a prompt upsampler”, and --no-pup sends your prompt as typed, so in that case supply a full scene caption written by another LLM using BFL’s system_prompts.json.

Common Problems

• Out of memory (OOM) errors: If you run out of GPU memory, turn off the prompt upsampler (pup, on by default in the reference CLI) with --no-pup, since pup adds about 28 GB of GPU memory (on an 80 GB GPU, per the model card). BFL says not to use the model without a prompt upsampler, so then supply a full scene caption written by another LLM using BFL’s system_prompts.json, because --no-pup sends your prompt as typed. Next, switch from BF16 (31 GB peak VRAM) to the FP8 (26 GB) or NVFP4 (25 GB) variants. If you have a smaller card, use the int8 ComfyUI files, which run on 8 GB GPUs such as the RTX 3070. See our VRAM page for details.
• Gated download errors: The FLUX 3 Klein weights are gated on Hugging Face. You must create or log in to a Hugging Face account and accept the FLUX Kommunity License v.1.0 on the model card before your local tools can download the files. For ComfyUI, the files are in the separately gated black-forest-labs/flux-3-image-klein-comfy repository: open it and accept the license agreement there too. See our download guide.
• ModularPipeline loading errors: If loading FLUX 3 Klein with ModularPipeline fails, use the Git install documented on the model card: pip install git+https://github.com/huggingface/diffusers.git.
• License limits for production or commercial products: The FLUX Kommunity License v.1.0 allows non-commercial use, plus commercial output generation by Qualifying Users (under US$5,000,000 in gross annualized revenue, including affiliates). However, BFL states that “Running Klein in a product needs a commercial license, which covers self-hosted deployment at scale.” Check our license guide before deploying the model in a business or product.

Wispvue (flux3klein.org) is an unofficial informational site and is not affiliated with or endorsed by Black Forest Labs.