Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, an open-weights text-to-image model that compresses the capabilities of the larger FLUX 3 Image family into 5.6 billion parameters1. The model handles text-to-image generation, pixel-level image editing, and multi-reference generation with up to 10 reference images.
Smaller footprint, consumer hardware
The release ships as a base model for fine-tuning and a step-distilled variant for low-latency inference, alongside FP8, INT8, and NVFP4 quantized builds developed with Nvidia Corp. support. Generator weights occupy 11.2 GB in BF16 and 5.6 GB in FP8; the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. Black Forest Labs said the reduced parameter count makes the model practical for local deployment, cheaper to adapt, and suited to on-device image editing on laptops.
On Nvidia Corp.'s GB300, the distilled model generates a 1-megapixel image in under 0.6 seconds in NVFP4.
Benchmarks and competitive positioning
In internal evaluations, Black Forest Labs reported that FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation, and image editing, rating above Ideogram V4.5 and FLUX.2 pro. Compared with FLUX.2 klein 9B, the new model gains 91 Elo in text-to-image and 70 Elo in image editing while using 38% fewer parameters. At 5.6B parameters it is the smallest open model in the company's comparison set, against 7B for Qwen-Image-2.1 and 12.9B for Krea 2. Black Forest Labs reported a median BFL API latency of 4.54 seconds for text-to-image and 7.18 seconds for editing at 1 MP.
Ecosystem and licensing
Day-one tooling support spans ComfyUI, Hugging Face Inc. Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni. Diffusers and Ostris's aitoolkit support LoRA training from launch, with serving a LoRA adding roughly 5% to generation time. Nvidia Corp.'s NeMo AutoModel ships training recipes that run directly from the released checkpoints. Base and distilled weights are posted on Hugging Face Inc., with reference inference code on GitHub and a Gradio demo for bounding-box control.
The weights ship under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue under $5 million, alongside research and non-commercial use. Larger organizations can deploy through self-hosted licensing or the BFL API, where pricing for the distilled model starts at $0.009 per 1,000 images.
ANALYSIS Packaging a competitive image-generation stack into 5.6 GB at FP8 lowers the hardware floor to mainstream gaming GPUs, a threshold that could shift fine-tuning and inference workloads off cloud APIs and onto local machines for cost-sensitive developers and small studios.