Black Forest Labs FLUX 3 Image Klein Open-Weight Image Model

Features

• ✓ Native 4K and pixel-perfect editing at 5.6B parameters: Black Forest Labs says FLUX 3 Image Klein shares core capabilities with the larger FLUX 3 Image model, including native 4K generation, pixel-perfect editing, and bounding boxes that control where elements are placed or confine an edit to a defined area
• ✓ One model for text-to-image, editing and multi-reference generation: BFL says it takes a prompt plus up to 10 reference images and returns one image, with text-to-image, single-reference edits and multi-reference compositions running through the same model, using Qwen3-VL-4B-Instruct as the text encoder
• ✓ It runs on consumer GPUs: BFL says the generator weights need 11.2GB in BF16 and 5.6GB in FP8, while the INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM. At 5.6B parameters it is a practical option for local deployment and for teams with data sovereignty requirements
• ✓ Concrete speed numbers: BFL says that with the step-distilled model and the INT8 build, a 1MP image takes about 4 seconds on an NVIDIA GeForce RTX 5090 and under 9 seconds on an RTX 4070 through ComfyUI, measured by NVIDIA
• ✓ Fine-tuning from day one: BFL says the base checkpoint is built for fine-tuning, with Diffusers and Ostris's AI Toolkit supporting LoRA training from day one and NVIDIA's NeMo AutoModel shipping recipes that work directly from the released checkpoints. The release includes a base model, a low-latency step-distilled model, and FP8, INT8 and NVFP4 builds developed with NVIDIA

How to Use

1. Get the weights from BFL's official blog or the black-forest-labs/flux-3-image-klein repository on Hugging Face: note this repository holds the distilled checkpoint for generation and image editing, and that fine-tuning, LoRA training and research should use FLUX 3 Image Klein Base
2. Pick a build for your VRAM: INT8 with 8GB, or FP8 (~5.6GB) and BF16 (~11.2GB) if you want more fidelity; BFL says the INT8 build runs in either ComfyUI or Diffusers
3. Start with text-to-image then move to editing: generate from a prompt first, then try single-reference edits that carry a subject or style from the reference, and multi-reference compositions with up to 10 inputs, and finally use bounding boxes to place elements or confine an edit to a region
4. Check the licence before commercial use: BFL says the release is under the FLUX Kommunity License, and third-party coverage reports free commercial use under $5 million in annual revenue with a separate arrangement above that threshold; read the full official licence text before shipping anything

FAQ

Official Source