Overview

FLUX 3 Klein is the short name for FLUX 3 Image Klein, an image model from Black Forest Labs. The model card says the 5.6B count is the image transformer only, and the model does text-to-image, single-image editing, and multi-reference editing with up to 10 images, plus bounding-box layout and in-image text, at up to 4K. The Black Forest Labs blog is dated 2026-10-15. That date is the public release used here.

ComfyUI and community posts also write the name as FLUX.3 [klein] and FLUX 3 Klein. The Hugging Face repository black-forest-labs/flux-3-image-klein was created on 2026-10-13 at 10:58 UTC. The blog date, not the repository creation time, is the public date.

The text encoder is not inside the 5.6B label. The model card names Qwen3-VL-4B-Instruct, under Apache-2.0. No file size for that encoder is in the verified table.

Klein vs Klein Base

There are two checkpoints.

Klein is distilled. The official template default is 16 steps. The model card describes it as 16 steps with no guidance, and the ComfyUI tutorial describes the distilled template as 16 steps at CFG 1, adjustable from 4 to 16. It is the checkpoint behind the Black Forest Labs API. The API model id is flux-3-image-klein. The blog says: “Pricing starts at $0.009 per 1K image.”

Klein Base is undistilled. It is 50 steps, and it is the checkpoint for fine-tuning and LoRA. The official template default is CFG 3.5, which matches the model card and the Black Forest Labs docs. The GitHub README lists 4.

Use Klein for ordinary generation and editing. Use Klein Base when the work is a fine-tune or a LoRA, at the default CFG of 3.5.

Text encoder and prompt upsampler

The text encoder and the prompt upsampler are not the 5.6B image transformer.

Qwen3-VL-4B-Instruct is the text encoder, Apache-2.0. The prompt upsampler is a different model. The model card says: “Do not use the model without a prompt upsampler.”

The official upsampler is flux-3-image-klein-pup. It is a 9B vision-language model fine-tuned from Qwen3.5-9B, also Apache-2.0. The BF16 file is 18.8 GB. That is the file size. On an 80 GB card that official BF16 upsampler adds about 28 GB while it runs. Our own test used the w4a8 upsampler (6.7 GB file) and added about 5.8 GB of peak VRAM, not 28 GB. The 18.8 GB and the 28 GB are both official BF16 figures, and they measure different things.

Any LLM may be used with the official system_prompts.json instead of that 9B file. The prompt still has to be expanded. Diffusers does not run the upsampler for you. The model card says to load it yourself with Transformers. The official CLI can run the upsampler alone, python -m flux3.pup, and then generate with --no-pup when both do not fit in memory together.

The upsampler repository is not gated. The weight repositories are.

Features

The model card, the GitHub README, and the Black Forest Labs blog list the same capabilities.

It generates images from text. It edits a single image. It takes reference images, up to 10. It reads bounding-box layout written in the prompt. It renders text inside the image. The maximum resolution they name is 4K.

Two box formats are named, and they are not the same scale. One is a 0–1000 grid written as [top, left, bottom, right]. The GitHub CLI’s boxes.json uses 0–1 decimals. Use the format that matches the tool in front of you. A 0–1 box pasted into a 0–1000 field, or the reverse, is a different layout.

Nothing in the verified table is an independent image-quality score. Black Forest Labs does publish its own comparison with FLUX.2 [klein] 9B: 38% fewer parameters, plus 91 Elo on text-to-image and plus 70 Elo on editing. That is BFL’s internal evaluation. The comparison set they name is Qwen-Image 2.1 at 7B and Krea 2 at 12.9B, with those parameter counts described as each publisher’s own figure. This page does not turn that into a ranking.

Quantized builds

Precision changes the file you download and the peak the publisher measured. It does not change the 5.6B label.

Generator weights in the Black Forest Labs blog: BF16 11.2 GB, FP8 5.6 GB.

ComfyUI files in the Hugging Face ComfyUI repository: BF16 11.1 GB, int8 5.8 GB, NVFP4 5.3 GB. The 11.2 GB blog figure and the 11.1 GB ComfyUI figure are both BF16, from different file sets. Do not treat them as one download.

On a B200, at 1024×1024, with the upsampler off, the model card and the GitHub memory notes give these peaks: BF16 31 GB, FP8 26 GB, NVFP4 25 GB. The official extra cost of about 28 GB on an 80 GB card is for the BF16 upsampler, and it is not inside that table. Our w4a8 upsampler (6.7 GB file) added about 5.8 GB of peak VRAM instead.

NVFP4 needs a Blackwell GPU. The ComfyUI repository and the GitHub notes are the sources. The 5.3 GB NVFP4 file is not a build for an RTX 30-series or 40-series card.

The INT8 path is the one tied to small NVIDIA cards. The blog says the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. The ComfyUI repository says: “The int8 files run on 8 GB GPUs such as the RTX 3070.”

By file size, using byte sizes from the Hugging Face API, the int8 generator is 5.83 GB, the int8 text encoder is 4.86 GB, the w4a8 upsampler is 6.72 GB, and the VAE is 0.34 GB. Together that is about 17.8 GB, so the full stack with the upsampler cannot be resident on an 8 GB card. The official 8 GB figure must refer to running without the upsampler resident, or to offloading or sequential loading. Whether an 8 GB card actually runs it is not something we tested.

NVIDIA’s speed measurement, distilled plus INT8, in ComfyUI, at 1 megapixel, is about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. The Black Forest Labs blog and the ComfyUI blog report it. It is NVIDIA’s number, not a test from this site, and it is not a BF16 or 4K figure.

Where to get it

Weights for the distilled model are black-forest-labs/flux-3-image-klein on Hugging Face. Klein Base is black-forest-labs/flux-3-image-klein-base. The upsampler is black-forest-labs/flux-3-image-klein-pup. The ComfyUI files are black-forest-labs/flux-3-image-klein-comfy. Inference code is the GitHub repository black-forest-labs/flux-3-image-klein, under Apache-2.0.

The weight repositories are gated with an “auto” gate. Log in to Hugging Face, accept the terms, and access is granted automatically. The upsampler repository is not gated.

Diffusers support uses ModularPipeline, installed from the Diffusers git main branch, not from a pinned release named in the verified table. ComfyUI support is native from v0.41.0 on 2026-10-15, and the Klein template needs ComfyUI 0.41.0 or newer, with a 0.41.1 backport. Day-one support is also listed for SGLang, vLLM through vLLM-Omni, and TensorRT-LLM on the branch feat/flux3-image-klein.

The weights are under the FLUX Kommunity License v.1.0. The verified summary of a Qualifying User, from sections 1.e and 2.a, is the person plus affiliates with annual revenue under US$5,000,000. The license file itself was not pasted here, so this page does not put clause text in quotation marks.

Limitations

The model card has its own limitations list. That list was not copied into the verified fact table, so it is not restated here and not paraphrased from memory.

What the verified table does treat as limits:

You do not run the model without a prompt upsampler, unless you have already expanded the prompt another way the card allows. NVFP4 is a Blackwell build. Weight downloads are gated. The 8 GB line and the RTX 3090 line still disagree. The full stack with the upsampler is about 17.8 GB of files, and we have not tested an 8 GB card. No third-party image test, no Mac result, and no community GGUF quality claim is in the table, so none of those appear as facts on this page.