Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter model that creates and edits images, supports native 4K output and accepts up to 10 reference images. Downloadable weights and access through the BFL API are available today, with API pricing starting at $0.009 per 1K image. Black Forest Labs says the INT8 version can run through ComfyUI or Diffusers on NVIDIA GeForce RTX hardware with as little as 8GB of VRAM.
The weights use the FLUX Kommunity License, which permits commercial use without charge for individuals and businesses whose annual revenue falls below $5M, plus research and non-commercial use. Larger organizations seeking commercial deployment can use the API or obtain a self-hosted license. That revenue threshold puts a limit on free commercial deployment even though the weights are downloadable.
The release offers two model options: a base checkpoint intended for fine-tuning and a step distilled version intended to reduce generation latency. NVIDIA helped develop versions in NVFP4, FP8 and INT8 formats. Generator weights occupy 5.6 GB in FP8, compared with 11.2 GB in BF16.
With the distilled model using INT8 through ComfyUI, Black Forest Labs reports roughly four seconds for a 1MP image on an NVIDIA GeForce RTX 5090 and less than nine seconds on an RTX 4070. NVIDIA measured those GPU timings using 16 steps and generating one image per run; they exclude the network and queueing time included in API latency measurements.
For editing, bounding boxes let users specify where an element belongs or restrict changes to a selected region. The API serves the distilled version as flux-3-image-klein, with prompt upsampling and pixel restoration to retain unchanged image regions exactly. Developers can train LoRAs using Diffusers or Ostris's AI Toolkit, while NVIDIA's NeMo AutoModel provides recipes for the released checkpoints.
In its internal evaluations, Black Forest Labs places FLUX 3 Image Klein ahead of the other models tested for editing, reference-based generation and text-to-image output. Against FLUX.2 [klein] 9B, it reports gains of 70 Elo for editing and 91 Elo for text-to-image while using 38% fewer parameters. Those parameter comparisons count the image generators and exclude text encoders.
