Smaller than the model it replaces
Black Forest Labs has released FLUX 3 Image Klein, an open-weights image model with 5.6 billion parameters in the generator, excluding text encoders. The lab says it handles text-to-image, image editing and generation from up to ten reference images, with native 4K output, pixel-accurate editing, and bounding boxes for placing elements or confining an edit.
The comparison that matters is with the model it replaces. FLUX.2 [klein] had 9 billion parameters; the new one has 38% fewer and, on the lab’s own evaluations, scores 91 Elo points higher on text-to-image and 70 higher on image editing.
What it needs to run
The generator weights are 11.2GB in BF16 and 5.6GB in FP8. Black Forest Labs says an INT8 build runs on GeForce RTX cards with as little as 8GB of VRAM through ComfyUI or Diffusers, producing a one-megapixel image in about four seconds on an RTX 5090 and under nine on an RTX 4070. At the other end, the distilled model in NVFP4 on a GB300 is under 0.6 seconds for the same image, and the base model at 50 steps ranges from 1.7 seconds on a GB300 to 43.8 seconds on a DGX Spark.
FP8, INT8 and NVFP4 builds were developed with Nvidia. Day-one support covers ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training through the Ostris AI Toolkit and Nvidia NeMo AutoModel recipes.
The claims, and who made them
The lab says the model ranks first among open models in text-to-image, multi-reference generation and editing, ahead of larger open models, and that it rates above the closed Ideogram V4.5 and the lab’s own FLUX.2 [pro]. Those are internal evaluations published by the vendor. The comparison set also includes Qwen-Image 2.1 at 7 billion parameters and Krea 2 at 12.9 billion.
On the API, Black Forest Labs reports median end-to-end latency including queueing of 4.54 seconds for text-to-image and 7.18 for editing, and lists pricing from $0.009 for a 1K image. Weights for the base and distilled models are on Hugging Face.
Licence and safety
The weights come under what the lab calls the FLUX Kommunity License: free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use, with larger organisations going through self-hosted licensing or the API.
On safety, the lab says it filtered the pre-training dataset, did targeted fine-tuning afterwards, red-teamed the model internally, and had it evaluated by an external partner, Cinder, which it says reported substantially fewer vulnerabilities than other popular open-weights models. That external report is the one document worth reading in full when it appears.
