Black Forest Labs has released FLUX 3 Image Klein, a 5.6B-parameter open-weights model for text-to-image generation, image editing and multi-reference generation. Base and distilled weights are on Hugging Face under the FLUX Kommunity License. The distilled model is also in the BFL API as flux-3-image-klein.

Klein is the smaller sibling of FLUX 3 Image and the second open-weights release in the FLUX 3 family, after FLUX 3 Action. Black Forest Labs announced it on October 15, 2026, and says it was optimized for consumer GPUs. With the INT8 build it runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers.
What it is
- 5.6B parameters in the image generator. The text encoder (Qwen3-VL-4B) is separate.
- Text-to-image, image editing and multi-reference generation with up to 10 reference images.
- Native 4K generation.
- Bounding boxes that control where elements are placed or confine edits to a defined area.
- Two checkpoints: a base model for fine-tuning (50 steps) and a step-distilled model for fast generation (16 steps).
- FP8, INT8 and NVFP4 builds, developed with NVIDIA support.
How it compares
In Black Forest Labs' internal evaluations (Elo ratings), Klein ranks first in text-to-image, multi-reference generation and image editing among the open-weights models compared. It also rates above the two closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro].
| Model | Type | Generator parameters |
|---|---|---|
| FLUX 3 Image Klein | Open weights | 5.6B |
| Qwen-Image 2.1 | Open weights | 7B |
| FLUX.2 [klein] 9B | Open weights | 9B |
| Krea 2 | Open weights | 12.9B |
Parameter counts are as each publisher reports them, without text encoders. Krea 2 does not support image editing and is not in the editing comparison. Against FLUX.2 [klein] 9B, Klein gains 91 Elo in text-to-image and 70 Elo in image editing, with 38% fewer parameters.
Speed
NVIDIA measured generation time for a 1024×1024 image. RTX results are in ComfyUI.
| Setup | Time per image |
|---|---|
| Distilled + INT8, RTX 5090 | about 4 s |
| Distilled + INT8, RTX 4070 | under 9 s |
| Distilled, 16 steps, GB300 | 0.55 s |
Tools and fine-tuning
- Day-one inference support in ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni.
- LoRA training with Ostris's AI Toolkit and Diffusers, and recipes in NVIDIA NeMo AutoModel.
License
The weights use the FLUX Kommunity License v1.0. It allows research and non-commercial use, and free commercial use for individuals and companies with annual revenue under $5M (the model card counts affiliates' revenue). It is not blanket permission for commercial fine-tuning or distribution. Larger organizations need a commercial license or the API. The Hugging Face repos are gated: you accept the terms before downloading.
Sources
Did this land? Leave a reaction.
Discussion
0 commentsMembers can join the discussion
Become a member to comment and reply to other members.
Become a MemberNo comments yet
Be the first to start the discussion.