Flux 3 Image Klein: open weights, 5.6B parameters

FLUX 3 Image Klein is a 5.6 billion-parameter open-weights image generation model tuned for text-to-image synthesis, pixel-perfect image editing, and multi-reference generation with up to ten reference images. It supports native 4K outputs, bounding-box control to place or confine elements, and inherits multi-modal training signals from FLUX 3’s image, video and audio research to improve spatial detail and realism. The release includes a base checkpoint for fine-tuning and a step-distilled low-latency model, with FP8, INT8 and NVFP4 builds developed with NVIDIA. Weights are available under the FLUX Kommunity License (free commercial use for organizations under $5M revenue) and are usable through the BFL API, Hugging Face, ComfyUI, Diffusers and other tooling; LoRA training and NeMo recipes are provided for adaptation.

Internal evaluations position it ahead of larger open models and some closed models in text-to-image, multi-reference and editing tasks, with a reported +91 Elo improvement in text-to-image and +70 Elo in editing versus a 9B-parameter predecessor while using 38% fewer parameters. Generator weights occupy 11.2 GB in BF16 or 5.6 GB in FP8; INT8 builds run on consumer GPUs with as little as 8 GB VRAM. The distilled INT8 build produces a 1MP image in ~4 seconds on an RTX 5090 and <9 seconds on an RTX 4070 via ComfyUI; API latencies reported are ~4.54 s for text-to-image and ~7.18 s for editing. Development included dataset filtering, targeted fine-tuning, internal red teaming and external safety evaluation to reduce misuse risks.