FLUX 3 Image Klein: 5.6B open-weights image model, INT8 build for 8GB RTX

Black Forest Labs (BFL) released FLUX 3 Image Klein on October 15, 2026. It is a 5.6B-parameter open-weights image model. BFL counts the image generator only, not the text encoder. The model does text-to-image, image editing, multi-reference generation with up to 10 reference images, and bounding-box layout control. BFL claims native 4K generation. BFL lists a step-distilled model at 16 steps and a base model at 50 steps. It says the generator weights need 11.2 GB in BF16 and 5.6 GB in FP8, and that the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM. Day-one support is listed for ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni. The weights use the FLUX Kommunity License. BFL says it allows research and non-commercial use, and free commercial use for individuals and companies with annual revenue under $5M. Larger organisations can deploy commercially through BFL's self-hosted licensing or its API. BFL's Hugging Face card gives the licence as v1.0 and names the text encoder as Qwen3-VL-4B-Instruct. The repo is gated: you must agree to share your Hugging Face information. With the reference code on a B200 at 1024x1024 and the prompt upsampler off, peak memory is 31 GB (BF16), 26 GB (FP8) and 25 GB (NVFP4). The card says the upsampler adds about 28 GB on an 80 GB GPU. BFL's separate ComfyUI repo, also gated, says not to use the model without a prompt upsampler. Its workflows run the INT8 upsampler file by default. The upsampler files are 18.8 GB (BF16), 9.8 GB (INT8) and 6.7 GB (W4A8). The repo says the INT8 model files run on 8 GB GPUs such as the RTX 3070, and on GPUs without FP8 support (RTX 30 series and older). It says NVFP4 needs a Blackwell GPU.