Image model

How to run FLUX.3 Klein locally.

FLUX.3 Image Klein runs natively in ComfyUI 0.41.0 and newer with three files: a 5.6B model, a Qwen3-VL 4B text encoder and the Flux.2 VAE. The int8 files fit an 8 GB graphics card. An optional fourth file, BFL’s prompt upsampler, turns a short request into the long caption the model was trained on.

Updated 21 Oct 20268 min read

Maker
Black Forest Labs
Released
Oct 202613 to 15 October, Klein and Klein Base
Licence
FLUX Kommunity v1.0commercial use under US$5M revenue
Memory
8 GB and upint8. NVFP4 on RTX 50 series

Klein or Base.

FLUX.3 Image Klein is Black Forest Labs’ 5.6B image model, released between 13 and 15 October 2026. It works in the Flux.2 latent space with the Flux.2 VAE, reads prompts through Qwen3-VL 4B, and makes new images or edits from reference images (BFL says up to 10). It comes in two versions.

  • Klein is the distilled model and the one to use. 16 steps, no CFG or guidance.
  • Klein Base is undistilled, made for fine-tuning and training LoRAs. It takes 50 steps at CFG 4. The file names say base.

Both come as int8 (int8-convrot), bf16 and NVFP4. Pick int8 for an 8 GB card or a GPU without FP8, NVFP4 for an RTX 50 series card, and bf16 when you have the memory for the full-size file.

Files you need.

One model, one text encoder, one VAE. All of them are in BFL’s flux-3-image-klein-comfy repo. It’s gated: sign in, accept the FLUX Kommunity License, and access is granted automatically.

  • Model flux3-image-klein-int8-convrot.safetensors ComfyUI/models/diffusion_models/ or -bf16, 11.1 GB, or -nvfp4, 5.3 GB (RTX 50 only). Base: flux3-image-klein-base-…, same sizes
    5.8 GB Download
  • Text encoder qwen3vl_4b_int8_convrot.safetensors ComfyUI/models/text_encoders/ or qwen3vl_4b_bf16.safetensors, 8.9 GB
    4.9 GB Download
  • VAE flux2-vae.safetensors ComfyUI/models/vae/
    0.3 GB Download
  • Prompt upsampler qwen3.5_9b_flux3_prompt_upsampler_w4a8 ComfyUI/models/text_encoders/ optional. _int8_convrot is 9.7 GB, _bf16 18.8 GB
    6.7 GB Download

The VAE is the same one every Flux.2 model uses, so if you already run Flux.2 Klein you have it. The prompt upsampler is a Qwen3.5 9B model that expands a short request into the long, detailed caption FLUX.3 was trained on. BFL recommends always using it.

ComfyUI/models
models/
├── diffusion_models/
│   └── flux3-image-klein-int8-convrot.safetensors
├── text_encoders/
│   └── qwen3vl_4b_int8_convrot.safetensors
└── vae/
    └── flux2-vae.safetensors

What fits your computer.

Klein at 1024 × 1024. The text encoder runs first and ComfyUI moves it out of the way before sampling, so the model file sets the limit. BFL’s own figures cover the int8 and NVFP4 files; treat the other tiers as estimates from file sizes.

  • 8 GBFits

    int8 model (5.8 GB). BFL says the int8 files run on 8 GB cards such as the RTX 3070.

  • No FP8Fits

    RTX 30 series and older can’t compute in FP8. The int8 files run there too.

  • RTX 50Fits

    NVFP4 (5.3 GB), the smallest file. It needs a Blackwell card and won’t run on older GPUs.

  • 16 GB+Fits

    bf16 (11.1 GB) with room left at larger sizes. BFL recommends 2K for the best results.

Set it up.

  1. Update ComfyUI

    FLUX.3 is native from ComfyUI 0.41.0. Older versions don’t have the flux3 text encoder type. ComfyUI Desktop updates itself; in a manual install:

    Terminal, in the ComfyUI folder
    git pull
    pip install -r requirements.txt
  2. Download the three files

    Model, text encoder and VAE from the list above. Sign in to Hugging Face and accept the FLUX Kommunity License on the repo page first; approval is automatic.

  3. Put them in their folders

    Model in diffusion_models, text encoder in text_encoders, VAE in vae. The prompt upsampler, if you get it, also goes in text_encoders. Restart ComfyUI.

  4. Check the loaders

    Load Diffusion Model gets the FLUX.3 file. Load CLIP gets the Qwen3-VL 4B encoder with type flux3. Load VAE gets the Flux.2 VAE.

  5. Write a detailed prompt and run

    FLUX.3 was trained on long, detailed captions. Write one, or run your request through the prompt upsampler first. The first run loads everything and takes longer.

Settings that work.

Both versions sample with euler and the simple scheduler. BFL’s reference code uses a time shift of alpha 16; ComfyUI’s built-in value for FLUX.3 is an older one.

Klein

Steps
16
CFG
nonedistilled, no guidance
Sampler
euler
Scheduler
simple
Shift
alpha 16BFL reference
Size
1024 to 2KBFL recommends 2K
Text encoder
Load CLIP, flux3

Klein Base

Steps
50
CFG
4
Sampler
euler
Scheduler
simple

BFL says FLUX.3 Klein goes up to 4K and recommends 2K for the best results. 2K takes roughly five times as long as 1024 × 1024, so 1024 is the quicker place to try prompts.

How fast.

GPUModelSizeTime
RTX 3070 class, 8 GBKlein, 16 steps1024 px1 to 1.5 min
RTX 3070 class, 8 GBPrompt upsamplerper run+0.5 to 1.5 min
RTX 3070 class, 8 GBKlein Base, 50 steps with CFG1024 px4 to 5 min

Rough estimates, not measurements. At 2K, expect about five times as long.

When it goes wrong.

Load CLIP has no flux3 type
ComfyUI is older than 0.41.0. Update it.
The Hugging Face download is refused
The repo is gated. Sign in and accept the FLUX Kommunity License on the repo page; approval is automatic.
The NVFP4 file won’t run
NVFP4 needs an RTX 50 series (Blackwell) card. Use the int8 file on anything older.
Klein Base is very slow
It runs 50 steps with CFG, which takes several times as long as Klein. Use Klein for making images and keep Base for training.
Results look plain from a short prompt
FLUX.3 was trained on long, detailed captions. Write more, or run the prompt upsampler first.

Questions.

Which text encoder does FLUX.3 Klein use?

Qwen3-VL 4B, loaded with a single Load CLIP node set to type flux3. The int8 file is 4.9 GB and the bf16 file 8.9 GB. It doesn’t use the Qwen3 encoders of Flux.2 Klein.

Should I get Klein or Klein Base?

Klein for making images: it’s distilled, runs 16 steps and needs no CFG. Klein Base is undistilled, takes 50 steps at CFG 4 and is meant for fine-tuning and LoRA training.

Do I need the prompt upsampler?

No, but Black Forest Labs recommends it. It’s a Qwen3.5 9B model that expands a short request into the long, detailed caption FLUX.3 was trained on. Without it, FLUX.3 runs on your prompt as written.

Can I use FLUX.3 Klein commercially?

It’s under the FLUX Kommunity License v1.0. Non-commercial use is allowed, and qualifying users with under US$5M gross annual revenue may use the outputs commercially. Other commercial use needs a licence from Black Forest Labs.

Does FLUX.3 Klein run on an 8 GB graphics card?

Yes, with the int8 files. Black Forest Labs names 8 GB cards such as the RTX 3070, and the int8 files also run on GPUs without FP8, such as the RTX 30 series and older.

Sources: FLUX.3 Image Klein ComfyUI files, ComfyUI repository.

HEISS UI

FLUX.3 Klein, ready to try.

HEISS UI runs FLUX.3 Klein on the ComfyUI you already have. Drop the model file into diffusion_models and it’s recognized.

  • Klein or Base, told apart by the weights. Not by the file name. Each gets its own steps and CFG, and BFL’s time shift, with no setting to touch.
  • The prompt upsampler, word for word. When it’s installed, every run uses it with BFL’s system prompts and settings. A switch in Basics turns it off and sends your prompt exactly as written.
  • Missing parts, one click each. The text encoder, prompt upsampler and VAE download from the setup panel. When BFL’s gated download is refused, it falls back to the same file from an open copy.
  • Edit from your own pictures. Three reference slots in the composer, 2K in the size menu, and FLUX.3 LoRAs recognized.

The model file itself you download yourself. Needs ComfyUI 0.41.0 or newer; older versions show “Newer ComfyUI”. Free and open source. macOS, Windows and Linux. Runs on your ComfyUI.

HEISS UI with a gallery of generated images and the prompt composer at the bottom.