The licence is the decision that matters most. This page is for developers who want to generate images through an API and need to know whether they may ship the result in a product. FLUX.2 Klein is published as four checkpoints (model files) with different licences. The 4B model is open under Apache 2.0, and the 9B model is not. So can you use FLUX.2 Klein in a product, and how do you call it?

What is FLUX.2 Klein?

FLUX.2 Klein is a family of image models that Black Forest Labs released on 15 January 2026. One model handles text-to-image, image editing and multi-reference generation, where you pass several input images. Our image model list also has FLUX.1 Dev, the earlier generation, which we cover in a separate guide. The name comes from the German word for "small".

The family has four checkpoints. Two are distilled, which means a large model taught a smaller one to reach the same picture in fewer steps, so they need only four. The two base ones are not distilled: they suit further training on your own data (fine-tuning).

CheckpointSizeStepsLicence
Klein 4B4 billion parameters4Apache 2.0
Klein 9B9 billion parameters4FLUX Non-Commercial License
Klein 4B Base4 billion parameters50Apache 2.0
Klein 9B Base9 billion parametersnot statedFLUX Non-Commercial License

The 4B model needs about 13 GB of VRAM (GPU memory), so it runs on a consumer card such as an RTX 3090 or 4070.

How fast is FLUX.2 Klein?

Black Forest Labs reports images in under 0.5 seconds on modern hardware, measured on its own data-centre GPU. Latency on our API also includes queueing and the network, so time your own calls before you promise a response time.

Black Forest Labs also ships 8-bit and 4-bit versions of the weights that use less VRAM and run faster on RTX 5080 and 5090 cards. Its blog lists the exact figures.

On quality, Black Forest Labs says Klein matches or exceeds the image quality of the Qwen image model at a fraction of the latency and VRAM. That is a vendor comparison, not a test of ours.

How do you call FLUX.2 Klein on Lyceum?

Image generation is not part of our OpenAI-compatible chat API. It has its own routes under `/api/v2/external/images`, and you authenticate with the same API key from your dashboard. Use the base URL shown in your Lyceum dashboard. The examples below use the one in our image generation reference.

First list the image models. The list models route returns an id and a name for each model, so you can look the id up by name instead of hard-coding it.

import os
import requests

BASE = "https://api.lyceum.technology"
HEADERS = {"Authorization": f"Bearer {os.environ['LYCEUM_API_KEY']}"}

models = requests.get(f"{BASE}/api/v2/external/images/models", headers=HEADERS).json()
model_id = next(m["id"] for m in models if m["name"] == "FLUX.2 Klein")

Then send the prompt. The request has three fields: `model`, `prompt` and `aspect_ratio`, which defaults to `"1:1"`. The response has one field, `image_url`.

response = requests.post(
    f"{BASE}/api/v2/external/images/generations",
    headers=HEADERS,
    json={"model": model_id, "prompt": "a lighthouse at dusk, film photo", "aspect_ratio": "1:1"},
)
print(response.json()["image_url"])

No field takes an input image. The model can edit images, but the editing and multi-reference modes are not available through this route, so for editing you would run the weights yourself. For per-image prices and volume, talk to the engineers who run Lyceum.

Your app sends a prompt, a model id and an aspect ratio to the Lyceum image route, which runs FLUX.2 Klein and returns one image URL. The route takes no input image.
Our image route takes a prompt, a model id and an aspect ratio, and returns one image URL; it takes no input image.

Can you use FLUX.2 Klein in a product?

The licence depends on the size. The 4B model is released under Apache 2.0, and its model card says the weights are available for commercial use. The 9B model falls under the FLUX Non-Commercial License.

Our list returns the name "FLUX.2 Klein" and an id, nothing more. It does not say which size runs behind it. Ask us before you plan production use, and have the answer in writing. If you need the 4B weights on your own hardware, a Lyceum GPU virtual machine is billed per second.

FLUX.2 Klein 4B is released under Apache 2.0 and 9B under the FLUX Non-Commercial License; our model list does not state which size runs behind the Lyceum model.
The 4B checkpoint is Apache 2.0 and the 9B checkpoint is non-commercial, and our model list does not say which size you get.

What does FLUX.2 Klein get wrong?

The model card for the 4B model lists the limits. Text inside an image can be inaccurate or distorted, so read every generated sign, label and button before you ship it. Prompt following depends heavily on the prompting style, and the model is not meant to provide factual information.

Treat the first week as a prompt-testing week. Generate the same ten prompts at the sizes you need, and keep the ones that hold up.

Where does your data go?

Lyceum is a sovereign European AI inference provider for open-source models, and the GPUs run in European data centres. Our trust page lists the data-handling terms for your security review.

Ready to try it? Talk to the engineers who run Lyceum about image prices, volume and which FLUX.2 Klein you get.

Sources

  1. bfl.ai
  2. huggingface.co
  3. huggingface.co
  4. huggingface.co