Black Forest Labs published open weights for FLUX 3 Image Klein on October 15, 2026, positioning the 5.6 billion-parameter release as a compact member of its FLUX 3 image line that still covers text-to-image, multi-reference image-to-image, and bounding-box layout control. The company says the model shares core capabilities with the larger FLUX 3 Image stack, including native 4K generation and edits that can be confined to defined regions.
The package ships two main checkpoints. A step-distilled build targets low-latency generation (16 steps, no classifier-free guidance by default in Diffusers). A base build (50 steps with guidance) is intended for fine-tuning and LoRA training. Black Forest Labs also published FP8, INT8, and NVFP4 transformer builds developed with NVIDIA support so the same generator can fit smaller GPUs. Hugging Face Diffusers added day-one Modular Pipeline support, and ComfyUI, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni are listed among supported runtimes.
Licensing is central to the story. Weights land under the FLUX Kommunity License, which allows free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use. Larger organizations are pointed to self-hosted commercial licenses or the BFL API. On the API, the distilled endpoint is named flux-3-image-klein, with pricing stated from $0.009 per 1K image, including prompt upsampling and pixel restoration that aims to leave unedited regions unchanged.
Black Forest Labs reports internal evaluations in which Klein ranks first among compared open-weight models on text-to-image, multi-reference generation, and editing, and also rates above Ideogram V4.5 and FLUX.2 [pro] in those charts. Those figures are vendor-reported Elo-style comparisons, not independent third-party leaderboards. The company also highlights consumer-GPU practicality: BF16 generator weights around 11.2 GB, FP8 around 5.6 GB, and INT8 paths that can run on NVIDIA GeForce cards with as little as 8 GB of VRAM via ComfyUI or Diffusers.
Layout control is the distinctive product claim. Prompts can include JSON bounding boxes (normalized coordinates) so regions are placed or moved with src_bbox and tgt_bbox semantics. Up to ten reference images can condition a generation. Architecture notes from Diffusers describe a latent flow model with Qwen3-VL text encoding, separate then joint transformer stages, and per-token timestep modulation that marks references as clean context.
Diffusers documentation also notes that reference tokens join the same sequence as text and the image being generated, so a single 1024 by 1024 reference roughly doubles end-to-end time on an H100 in their measurements. That trade-off is relevant for builders choosing between API latency and local multi-reference editing.
Safety work is described as dataset filtering, post-training suppression of harmful behaviors, internal red teaming, and evaluation by external partner Cinder. FLUX 3 Image Klein is framed as the second open-weight drop in the FLUX 3 family after FLUX 3 Action. Independent coverage on Hugging Face Blog confirms Diffusers and Spaces availability the same day.