FLUX 3 Image Klein is out: generation, editing and 10 references in a 5.6B local model
16 octobre 2026 · Kris Wilhem
Black Forest Labs has just released FLUX 3 Image Klein, the compact, open-weights version of its new image generation.
And this time, it is not one model for generating and another for editing.
The same checkpoint can do text-to-image, editing, multi-reference composition and layout control through bounding boxes.
For us, that is probably the most interesting point of this release.
One model to generate and edit
FLUX 3 Image Klein accepts a prompt and up to 10 reference images.
It can be used to:
- generate an image from text;
- modify an existing image;
- combine several references;
- move, replace or recolour elements;
- control their position and scale with bounding boxes;
- generate text inside the image.
The model uses Qwen3-VL-4B-Instruct as its encoder for text and visual references.
The announced 5.6 billion parameters concern the image transformer; the Qwen encoder is separate.
Klein and Klein Base
The release includes two main versions.
FLUX 3 Image Klein
The distilled version is designed for inference.
It runs by default at 16 steps, without CFG, and it is also the checkpoint used behind the Black Forest Labs API.
FLUX 3 Image Klein Base
The Base version is not distilled.
It runs at around 50 steps with guidance and is intended first and foremost for:
- fine-tuning;
- LoRA training;
- research;
- custom adaptations.
Black Forest Labs provides from launch the tools needed to train LoRAs with Diffusers, AI Toolkit and NVIDIA NeMo AutoModel.
A model genuinely designed to run locally
This is probably the other important piece of news.
FLUX 3 Image Klein is supplied in several formats:
- BF16;
- FP8;
- INT8;
- NVFP4.
Black Forest Labs states that the INT8 version can run on NVIDIA GeForce cards with 8 GB of VRAM.
In the measurements published with NVIDIA, the distilled model produces a 1-megapixel image in about 8.9 seconds on an RTX 4070 via ComfyUI.
In other words, we are no longer looking only at a theoretical "open weights" model that needs a machine out of reach for most creators.
It is starting to enter the category of models that can genuinely be used on a local workstation.
ComfyUI supported from day one
Black Forest Labs directly provides a repackaged version for ComfyUI.
It includes in particular:
- the distilled transformer;
- the Base model;
- various quantisations;
- the text encoder;
- the autoencoder;
- the prompt upsampler.
The model is also supported from release by Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni.
For a studio already working with ComfyUI, this greatly shortens the distance between a model's announcement and its real integration into a pipeline.
The return of bounding boxes
FLUX 3 Image Klein also picks up a direction we follow very closely at Axiris: building the image with explicit zones and positions rather than leaving everything to a single text description.
Bounding boxes are expressed in the prompt on a grid from 0 to 1000.
They can be used to guide:
- position;
- size;
- movement;
- replacement;
- the recolouring of an element.
One must, however, distinguish spatial guidance from a hard mask.
In the open-weights version, Black Forest Labs specifies that the boxes guide placement and scale but do not on their own constitute a pixel-perfect constraint comparable to a locked mask.
On the API side, BFL additionally offers a pixel restoration mechanism meant to preserve the unmodified regions exactly.
This nuance matters for real editing workflows.
Editing finally becomes a native capability
This is probably where this release is most interesting.
For a long time we separated:
```text GENERATION │ ▼ T2I model
EDITING │ ▼ another model / inpainting / dedicated workflow ```
With FLUX 3 Image Klein, Black Forest Labs pushes further towards:
```text FLUX 3 IMAGE KLEIN │ ┌──────────────┼──────────────┐ ▼ ▼ ▼ generation editing references │ │ │ └──────────────┼──────────────┘ ▼ same pipeline ```
This is an important evolution for production.
A single engine can understand the scene, receive several references and then intervene on the image without a complete change of system.
Why this release interests us at Axiris
At Axiris, our goal is no longer to look for "the model that does everything".
We are progressively building an infrastructure able to select the best engine for each operation.
FLUX 3 Image Klein becomes immediately interesting in this context because it brings together several criteria we are looking for:
- local execution;
- reasonable size;
- generation;
- editing;
- multi-reference;
- spatial control;
- LoRA;
- ComfyUI;
- quantisation suited to consumer GPUs.
It can therefore be evaluated not only as an image generator, but as a complete building block for local visual production.
And as always with us, the next step is not to repeat the maker's benchmarks.
It is to put it into our workflows and watch what it really does.
Licence
FLUX 3 Image Klein is published under the FLUX Kommunity License.
Black Forest Labs states that this licence allows free commercial use for individuals and for companies with less than 5 million dollars in annual revenue, subject to the terms of the licence.
Organisations above that threshold must use a suitable commercial licence or the BFL API.
Sources
- Black Forest Labs: FLUX 3 Image Klein, 15 October 2026
- Black Forest Labs: FLUX 3 documentation and models
- Hugging Face: black-forest-labs/flux-3-image-klein
- Hugging Face: black-forest-labs/flux-3-image-klein-base
- Hugging Face: FLUX 3 Image Klein for ComfyUI