FLUX 3 Image Klein: which version to download, and what it really needs

Which safetensors file should you download?

Black Forest Labs publishes the distilled model in four precisions. Sizes are as listed on Hugging Face:

File names in the flux-3-image-klein-comfy repository: flux3_image_klein_int8_convrot, flux3_image_klein_nvfp4 and flux3_image_klein_bf16 (all .safetensors). FP8 sits in the main flux-3-image-klein repository as variants/fp8r. Black Forest Labs says the INT8 build runs in as little as 8 GB of VRAM.

You also need the text encoder (Qwen3-VL-4B: 4.86 GB in INT8 or 8.88 GB in BF16) and the VAE (0.34 GB). An INT8 setup is about 11 GB of downloads before any prompt upsampler.

GGUF? Black Forest Labs has not published GGUF files in its repositories. Community GGUF and Mac MLX conversions appeared on Hugging Face within a day; we have not tested them, so treat them as unofficial.

The files are gated: log in to Hugging Face and accept the conditions before they unlock. Download only from the black-forest-labs organisation; reuploads appear fast after a big release.

How to set it up in ComfyUI

1. Update ComfyUI to the latest version.
2. Download the files from the flux-3-image-klein-comfy repository. It uses the same folder names as ComfyUI: put the model in models/diffusion_models, the Qwen3-VL-4B text encoder in models/text_encoders and the VAE in models/vae.
3. Open the FLUX 3 Image Klein template from ComfyUI's templates.
4. If you build the graph yourself, it uses the standard loaders: Load Diffusion Model for the checkpoint, Load CLIP with the type set to flux3 for the text encoder, and Load VAE.

No API key is needed to run it locally, according to Comfy's own launch page.

Klein or Klein Base?

If you just want images, take Klein. Base is three times the steps and exists to be trained on: Diffusers and Ostris's AI Toolkit support LoRA training on it from day one.

Speed, as measured by NVIDIA in ComfyUI with the distilled model and the INT8 build: a 1 megapixel image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070.

How much VRAM do you really need?

The "8 GB" headline and the model card describe two different setups:

• ComfyUI or Diffusers with the INT8 build: as little as 8 GB of VRAM on a GeForce RTX GPU, according to Black Forest Labs.
• Black Forest Labs' reference code: peak memory of 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4 (1024x1024 on a B200, prompt upsampler off). Their own upsampler adds about 28 GB on top.

So on a consumer card, stay in ComfyUI with a quantized build. The model card also warns that the file size of a quantized package is not a measure of total GPU memory use.

The prompt upsampler catch

The model card is direct about it: "Do not use the model without a prompt upsampler." Klein was trained on full scene captions, so a short prompt sent as typed gives generic output, and a one-line edit like "make it blue" can reframe the whole shot.

You have three options:

• The official upsampler, flux-3-image-klein-pup (18.8 GB). The reference code loads it by default.
• The ComfyUI upsampler builds in the comfy repository: 9.75 GB in INT8 or 6.72 GB in W4A8.
• Any LLM you already use, given the system prompt from the official system_prompts.json, then paste its caption in as the prompt.

Diffusers does not run an upsampler for you, so plain Diffusers scripts need one of these in front.

Can you use it commercially?

The weights use the FLUX Kommunity License v1.0. In plain terms:

• Non-commercial and research use is allowed.
• Commercial use of the images you generate is allowed for "Qualifying Users" with less than US$5M in gross annualized revenue, including affiliates.
• It is not blanket permission for commercial fine-tuning or for distributing the model. Running Klein inside a product at scale needs a commercial license from Black Forest Labs, or you can use their API.
• The Qwen3-VL-4B text encoder is under Apache-2.0.

The license also bans uses such as surveillance, biometric processing, non-consensual intimate images and training a competing image model on its outputs. We are not lawyers: read the full license on the model card before you build a business on it.

FLUX 3 Image Klein vs FLUX.2 klein 4B and 9B

FLUX.2 klein came in two sizes, 4B and 9B. FLUX 3 Image Klein comes in one size, 5.6B, so there is no "4B or 9B" choice this time: the choice is the precision in the table above.

• Smaller: 5.6B parameters against 9B for FLUX.2 [klein] 9B, 38% fewer.
• Better, by BFL's own tests: +91 Elo in text-to-image and +70 in image editing over FLUX.2 [klein] 9B. Independent benchmarks were not out when we checked.
• Bounding boxes: place elements with boxes on a 0 to 1000 grid, written as [top, left, bottom, right]. Boxes guide placement and scale; they are not a hard mask, so edits can still spill outside them.
• Native 4K: BFL says Klein keeps native 4K generation from the larger FLUX 3 Image model.

No GPU? Try it online first

• Official Hugging Face Space: free demo in the browser.
• BFL Playground and API: the distilled model as flux-3-image-klein, paid per image, with prompt upsampling built in. Check the BFL pricing page for the current rate.
• Comfy Cloud: ComfyUI on cloud GPUs; it offered 5 free runs without a card when we checked, then paid plans.

Testing online first tells you whether Klein suits your work before you spend an evening on an 11 GB download.

How we wrote this page: every figure comes from Black Forest Labs, its Hugging Face model cards or NVIDIA, linked below. We did not benchmark the model ourselves. More on how we check facts.