FLUX 3 Image Klein Official Model Overview and Local Setup

Yes—FLUX 3 Image Klein is the official Black Forest Labs model announced on October 15, 2026. It is a 5.6-billion-parameter open-weights model for text-to-image generation, image editing, and multi-reference creation; Hugging Face uses the IDs black-forest-labs/flux-3-image-klein and black-forest-labs/flux-3-image-klein-base. For local use, Black Forest Labs says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM.

What Is FLUX 3 Image Klein?

Some Diffusers documentation styles the name FLUX.3 Image Klein, while Black Forest Labs’ announcement uses FLUX 3 Image Klein. They refer to the same release; the Hugging Face repository IDs use lowercase, hyphenated names.

Klein provides text-to-image generation, image editing, and multi-reference generation with up to 10 reference images. The model also supports native 4K generation and bounding boxes for controlling element placement or confining edits to selected areas.

Black Forest Labs released two primary checkpoints:

Checkpoint	Intended use	Published defaults
black-forest-labs/flux-3-image-klein	Step-distilled model for fast inference	16 steps, no classifier-free guidance
black-forest-labs/flux-3-image-klein-base	Base model for fine-tuning and base sampling workflows	50 steps with classifier-free guidance

The distilled model produces a 1024 × 1024 image by default in the documented Diffusers workflow. The 5.6B parameter count covers the image generator; Black Forest Labs says it does not include the separate text encoder.

Official Download and Documentation Links

To obtain the model files and documentation, start with the Official FLUX 3 Image Klein announcement from Black Forest Labs to confirm the release, capabilities, supported formats, benchmarks, and license. For the Distilled Hugging Face model card, users should download the standard inference checkpoint and review its files, pipeline configuration, examples, and current license. Additionally, the Base Hugging Face model card serves as the starting point for LoRA training and fine-tuning. For hosted generation, consult the BFL API documentation to generate images without managing local weights or GPU infrastructure. Use the black-forest-labs owner path and official announcement rather than similarly named community uploads.

The release includes BF16, FP8, INT8, and NVFP4 options. However, availability is not identical across every loader: the Diffusers walkthrough specifically documents single-file loading for NVFP4 and FP8. Follow the exact filename and loader instructions on the selected model card rather than assuming all quantized files are interchangeable.

Model Size and VRAM Requirements

The following table details the published file sizes and runtime requirements for the model's BF16 and FP8 builds, alongside the minimum VRAM for INT8 configurations.

Measure	Published figure
BF16 generator weights	11.2 GB
FP8 generator weights	5.6 GB
Minimum VRAM stated for INT8	8 GB on supported NVIDIA GeForce RTX GPUs
Distilled INT8 speed on an RTX 5090	About 4 seconds for a 1 MP image
Distilled INT8 speed on an RTX 4070	Under 9 seconds for a 1 MP image

These measurements were reported by Black Forest Labs from NVIDIA testing with the distilled model at 16 steps in ComfyUI. They are not guarantees for every prompt, resolution, framework, or system.

The weight figures do not represent the complete installation size. Klein also uses a Qwen3-VL text encoder and other pipeline components, so disk space, system RAM, and runtime memory can be higher than the transformer file alone. Black Forest Labs has not published one exact total size for every quantized package.

The official 8 GB claim is specifically for NVIDIA GeForce RTX hardware. Do not assume the same performance or compatibility for AMD, Apple, or older NVIDIA GPUs without checking the relevant framework documentation.

How to Install It With Diffusers

Use current PyTorch, Diffusers, and Transformers releases. The documented dependency command is:

pip install -U diffusers transformers

The following example loads the standard distilled repository and generates an image:

import torch
from diffusers import ModularPipeline

pipe = ModularPipeline.from_pretrained(
    "black-forest-labs/flux-3-image-klein"
)
pipe.load_components(dtype=torch.bfloat16)
pipe.to("cuda")

image = pipe(
    prompt="A red fox standing in a snowy forest at golden hour",
    output="images",
)[0]

image.save("fox.png")

This sample uses BF16 and is not an 8 GB VRAM recipe. To run on an 8 GB card, use the INT8 build identified in the current official model card and load it through the documented ComfyUI or Diffusers path.

For local deployment, start by choosing the distilled or base checkpoint based on the specific requirements. Use the distilled checkpoint for ordinary text-to-image and editing work. Alternatively, select the base checkpoint when training a LoRA or preserving the published base sampling setup. Next, choose BF16 for maximum straightforward compatibility if hardware allows, or use FP8, INT8, or NVFP4 when lower memory use is more important. Finally, keep the initial test resolution at 1024 × 1024. Native 4K output is supported, although it requires substantially more memory and will be slower than the published 1 MP benchmarks.

Black Forest Labs also reports support for LoRA training through Diffusers, Ostris’s AI Toolkit, and NVIDIA NeMo AutoModel.

Prompts, References, and Layout Control

For text-to-image prompts, describe the subject, setting, lighting, composition, and visual style in complete sentences. The distilled model’s documented defaults are 16 steps without classifier-free guidance.

For multi-reference workflows, pass the images in a list and identify them in the prompt as <ref_image_0>, <ref_image_1>, and so on. Klein supports up to 10 reference images.

Layout regions are appended as a JSON list. Each region contains an id, a bbox, and a desc. Bounding boxes use:

[y_min, x_min, y_max, x_max]

Coordinates are normalized to a 0–1000 scale, with the vertical coordinate first. The same normalized box can therefore be reused at different output sizes.

License and Commercial Use

Klein is distributed under the FLUX Kommunity License, not as an unrestricted public-domain model. Launch coverage reports free commercial use for individuals and organizations below a $5 million annual-revenue threshold; larger organizations may need separate self-hosted licensing terms.

The current license file in the official Hugging Face repository is the controlling source. A free weight download also does not include a local GPU, and using the BFL API is a separate hosted service with its own pricing and terms.

FAQ

Is “FLUX 3 Image Klein” the correct official name?

Yes. Black Forest Labs’ announcement uses FLUX 3 Image Klein. Some Diffusers materials write FLUX.3 Image Klein, while Hugging Face repositories use black-forest-labs/flux-3-image-klein and black-forest-labs/flux-3-image-klein-base.

Is it the same as FLUX.2 [klein]?

No. FLUX 3 Image Klein is a new 5.6B-parameter release. Black Forest Labs says it has 38% fewer parameters than the 9B FLUX.2 [klein] model and adds multi-reference generation and bounding-box layout control.

How large is the model download?

Black Forest Labs reports 11.2 GB for the BF16 generator weights and 5.6 GB for FP8. Those are not complete installation sizes because the text encoder and other pipeline components require additional storage.

Can FLUX 3 Image Klein run with 8 GB VRAM?

Yes, Black Forest Labs says the INT8 version can run on supported NVIDIA GeForce RTX GPUs with as little as 8 GB VRAM. Use the quantized build rather than loading the standard BF16 pipeline unchanged.

Where can I download it for free?

Download the weights from the official Black Forest Labs collection on Hugging Face. There is no stated weight-download fee, but commercial use remains subject to the current FLUX Kommunity License.