FLUX.2 Klein Guide: Choose 4B, 9B, or Base Model Checkpoints

FLUX.2 Klein is Black Forest Labs' compact family for generating images and editing from references. It includes downloadable 4B and 9B models, with distilled and Base options, plus hosted access to distilled models. 1, 2

What are the key facts about FLUX.2 Klein?

When should you choose distilled Klein or Base?

Klein puts text-to-image generation and image editing in the same model family. Its compact options let you choose an image transformer separately from the much larger FLUX.2 dev checkpoint. 1, 5

Two decisions define the original Klein lineup: model size and distillation. The 4B and 9B labels identify different transformer sizes; Base identifies an undistilled checkpoint intended for more flexible development and training. 1

The distilled releases target short inference schedules. BFL's 4B model card provides a generation example using four steps, while the Base models retain the undistilled starting point for customization. 1, 3

For an interactive application, begin by evaluating the distilled checkpoint. For a training project, start with BFL's Base guidance. These are recommendations based on the purposes BFL assigns to the releases. 1

Do not make your first decision from a model filename alone. Write down whether your task is generation, image editing, or training, and whether you need to operate the model yourself.

For example, a tool that previews packaging concepts needs a usable turnaround and recognizable products. A project exploring a visual style may care more about a suitable training starting point than rapid single-image previews.

What are the license and hardware limits of Klein?

The 4B and 9B releases have different licenses. Apache 2.0 applies to the 4B models; the downloadable 9B models use BFL's non-commercial license. Check the specific model card before choosing a deployment route. 1, 3, 4

The parameter count describes the image transformer, not the entire application's memory footprint. BFL separately identifies an 8B Qwen3 text embedder for the 9B model in its launch post. 1

Hardware figures depend on configuration. The 4B card quotes about 13GB VRAM, while the inference repository quotes a lower figure without an equivalent setup explanation. Plan from the actual workflow you will run. 3, 5

That discrepancy is a reason to test a known example before committing to hardware. Record precision and offloading choices with your measurements so that another person can understand what the result describes.

The 4B model card lists inaccurate rendered text, prompt-following failures, and bias among its limitations. It also says the model is not intended to provide factual information. 3

For a practical trial, use a few representative assets and decide acceptance criteria in advance. Check a product's outline, a label's spelling, and the requested background independently instead of using one overall quality impression.

How do you run the FLUX.2 Klein 4B example?

1. Select a checkpoint deliberately. Start with black-forest-labs/FLUX.2-klein-4B if you want to reproduce the distilled 4B model-card example. Base 4B is a different checkpoint with a different intended use. 1, 3

2. Use a Python environment with a compatible CUDA-enabled PyTorch installation. Follow the model card's Diffusers installation instructions and check that Flux2KleinPipeline is available. 3

3. Load the pipeline and run the card's settings with your own prompt. This example enables model CPU offloading as shown in the official code; it does not establish a universal hardware minimum. 3
   

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained(
    "black-forest-labs/FLUX.2-klein-4B",
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
image = pipe(
    prompt="A blue teapot on a linen tablecloth, soft window light",
    height=1024,
    width=1024,
    guidance_scale=1.0,
    num_inference_steps=4,
    generator=torch.Generator(device="cuda").manual_seed(0),
).images[0]
image.save("klein-teapot.png")

1. Save the checkpoint name, prompt, dimensions, and settings with your output. Use the same task when you evaluate another variant so that you can explain what changed.

2. For a visual workflow, follow the ComfyUI implementation linked from BFL's model card. Begin with a workflow made for Klein and verify all requested model components before adding custom nodes. 3

The ComfyUI guide introduces the graph concepts. Keep model installation questions tied to the specific Klein workflow you select.

1. If you choose the API, use the documented Klein endpoint for your selected size. BFL distinguishes the hosted distilled models from Base checkpoints, which are provided as weights for local development. 2

Test ordinary language prompts before creating a complicated prompt generator. Describe the subject, material, action, setting, and light clearly, then check which requirements the model actually follows.

For a second pass, change only the requirement that failed. If the object is correct but the background is unsuitable, keep the object description stable while revising the setting.

How do Klein 4B, 9B, and Base compare?

These categories describe product choices, not a universal quality ranking. Pick the option that meets the job's output, deployment, and licensing needs, then validate it with representative examples.

The sibling FLUX.2 family overview places Klein alongside dev and the hosted models. Use it when the larger decision is local operation versus managed access.

What model sizes does FLUX.2 Klein offer?

FLUX.2 Klein's original lineup contains 4B and 9B image transformers. Each size has a distilled release and an undistilled Base checkpoint. 1

Are all Klein weights Apache 2.0?

FLUX.2 Klein 4B and Base 4B use Apache 2.0. Klein 9B and Base 9B weights use BFL's non-commercial license; check the chosen model card's terms. 1, 3, 4

Should I use Base for ordinary image generation?

BFL positions distilled FLUX.2 Klein for interactive generation and Klein Base for customization and training. Start with the distilled example for an ordinary generation trial. 1

Does 4B tell me how much GPU memory I need?

FLUX.2 Klein 4B names the image transformer's size, not the pipeline's VRAM requirement. The model-card example uses CPU offloading; measure the configuration you intend to run. 1, 3

Sources

• Klein release and variant definitions
• BFL model and access overview
• Official Klein 4B model card and example
• Official Klein 9B model card
• Official FLUX.2 inference repository

Related guides on PromptZone

• Realistic Photos with FLUX
• Best SDXL Models in 2026