CivArchive
    Flux.2 Klein - 4B
    NSFW
    Preview 119138979
    Preview 119138982
    Preview 119138999

    The FLUX.2 [klein] model family are our fastest image models to date. FLUX.2 [klein] unifies generation and editing in a single compact architecture, delivering state-of-the-art quality with end-to-end inference in as low as under a second. Built for applications that require real-time image generation without sacrificing quality.

    Originally Posted;

    Description

    FLUX.2-klein-4B

    FAQ

    Comments (25)

    Subseven8888Feb 2, 2026· 4 reactions
    CivitAI

    Can it be utilized in Forge Web UI?

    msuxFeb 2, 2026· 3 reactions

    Yes, in Forge Neo: https://github.com/Haoming02/sd-webui-forge-classic

    liutyiFeb 4, 2026· 1 reaction

    You can try SD.Next it works there

    zGenMediaFeb 5, 2026
    CivitAI

    I know this may sound dumb but I could not find an answer. Can this be used with older Flux loras? Flux1

    msuxFeb 5, 2026· 1 reaction

    No, they are very different models.

    sekosekoFeb 11, 2026· 1 reaction

    no you cant even use 4b models with 9b or vice versa.

    ZootAllures9111Feb 7, 2026· 9 reactions
    CivitAI

    Just reposting this comment I originally posted on the Flux.2 page before Klein got moved here, if it's useful to anyone:

    GGUFs for every version of Klein and both text encoders (Klein 9Bs need Qwen 8B, Klein 4Bs need Qwen 4B) are here BTW:

    https://huggingface.co/unsloth/Qwen3-4B-GGUF
    https://huggingface.co/unsloth/Qwen3-8B-GGUF

    https://huggingface.co/unsloth/FLUX.2-klein-4B-GGUF

    https://huggingface.co/unsloth/FLUX.2-klein-9B-GGUF

    https://huggingface.co/unsloth/FLUX.2-klein-base-4B-GGUF

    https://huggingface.co/unsloth/FLUX.2-klein-base-9B-GGUF

    UnsensualBurgerFeb 8, 2026
    CivitAI

    The image edit abilities are useful and comparable if not better in some areas as other models like QWEN but T2I ability is definitely worse for me.

    promptkidFeb 15, 2026· 3 reactions
    CivitAI

    Adherence to the prompt, variety of styles, not extremely resource-intensive (I'm running this on a 2070 in a tolerable way; Q8 generates in <10 seconds), i2i so easy it seems unreal.

    I think generative AI can't go much further, at least not for free...

    VV24Feb 15, 2026· 2 reactions
    CivitAI

    Please enable negative prompts in the generator. Since 9B/4B Base are undistilled models, they should be able to handle them as far as I understand. Negative prompts seem to work fine in ComfyUI for example.

    khajiirahFeb 19, 2026· 7 reactions
    CivitAI

    Has somebody a link for a working text encoder? qwen_3_4b.safetensors is not working. The Dimensions are not the same and KSampler generation fails. Thanks.

    carcassi75588Feb 23, 2026

    use qwen3_8b

    jccovert28Feb 26, 2026

    @carcassi75588 I used qwen_3_8b_fp8mixed with Flux 2 Klein 9b and I keep getting this error "KSampler mat1 and mat2 shapes cannot be multiplied (512x4096 and 12288x4096) TIPS: If you have any "Load CLIP" or "*CLIP Loader" nodes in your workflow connected to this sampler node make sure the correct file(s) and type is selected."

    robsonrds29700Feb 27, 2026
    CivitAI

    I can use this model with a 1080ti, but I use a python script to generate images dom SDXL and SD1.5, I like to adapt to use this type of model too already flux I could never adapt the script and my VGA little memory, I need what necessarily to run this, sorry for the ignorance...

    nerfmeMar 4, 2026
    CivitAI

    The 'edit' function reliably fails when using 9b and 9b-base on-site.

    theallyMar 4, 2026· 1 reaction

    Can you expand on this a little, please? I've just tested the Edit function, and was able to change specific parts of an image via text. None of my generations failed. If you see continued failures, please reach out to https://civitai.com/bugs with details.

    nerfmeMar 5, 2026

    Specifically it rejected the attempt, if that's the expression, it took buzz, then refunded buzz, over and over again. I'll try again sometime to see if it was just a server issue.

    nerfmeMar 5, 2026

    Tried again just now, I got 4/9 through. Thanks. It comes out great.

    elevendrMar 13, 2026
    CivitAI

    I have a computer with a 6GB GPU and a 16GB RAM. I am planning to use a low vram optimization workflow and a quantized variant instead of a full version. I want a model that does not sacrifice quality. I would want to know if I should get a quantized model of 9B Base, 9B, 4B Base, or 4B? Would like to know.

    StillbornMar 13, 2026· 1 reaction

    I have a RTX 4050 laptop with a 6gb VRam GPU (You can see my PC full configuration on my profile page) and i use the Flux.2 Klein 9B model with Loras without problems, but i have 32gb of Ram. I think you can run a 4b or 9b model too.

    lonecatone23Mar 16, 2026· 1 reaction

    You'll want to run 4b. Here's a workflow optimized for it.


    https://civitai.com/models/2414019/klein4b-essentials

    chrisss1Mar 24, 2026

    @Stillborn What text encoder and vae are you using for flux2 klein 9b?

    lonecatone23Mar 24, 2026· 1 reaction

    @chrisss1 It's listed in the notes, but it is qwen_3_8b and Flux2 vae

    chrisss1Mar 24, 2026

    @lonecatone23 Alright thanks

    StillbornMar 25, 2026· 1 reaction

    @chrisss1 If i remember correctly, I'm using the traditional Black Forest Labs models. You can find them on Hugging Face or here too. And, If you need any tip or anything, just send a message.

    Checkpoint
    Flux.2 Klein 4B
    by CivitaiOfficial
    Download (Beta) View on CivitAI
    base model

    Details

    Downloads
    2,193
    Platform
    CivitAI
    Platform Status
    Available
    Created
    1/28/2026
    Updated
    6/18/2026
    Deleted
    -

    Files

    flux2Klein_4b.safetensors

    Size:
    7.22 GB
    SHA256:
    ec3d4e733a771f61c052fb4856c48b336c55eaf2c65487c2a1faeb9bbda7a343

    Mirrors

    HuggingFace (45 mirrors)
    CivitAI (1 mirrors)
    ModelScope (1 mirrors)
    ModelScope CN (1 mirrors)