When Black Forest Labs shipped FLUX 3 Image in early October, the announcement noted open weights were coming in the following weeks. On October 15th they arrived — but not as the open-weights version of FLUX 3 Image. Instead, Black Forest Labs released FLUX 3 Image Klein: a distinct 5.6B model designed from the ground up to be compact, fast, and consumer-accessible, while claiming to match or exceed the larger model in quality.
The benchmark numbers are vendor-reported and should be read as such, but they’re interesting: 91 Elo points above FLUX.2 [klein] 9B in text-to-image, with 38% fewer parameters. Klein also claims to edge out the closed FLUX 3 Image in several tasks. If those numbers hold under independent evaluation, it would be one of the cleaner demonstrations that distillation and architecture work can deliver a better model at smaller scale, not just a cheaper one.
Klein inherits the full FLUX 3 Image feature set: bounding box layout control for placing elements at specific positions in the frame, multi-reference image conditioning (up to 10 reference images), pixel-perfect editing, and native 4K output. Nothing was stripped for the compact version. What changed is the weight count and the memory footprint. BF16 weights need 11.2GB; FP8 quantization brings that to 5.6GB. INT8 builds run on RTX GPUs with 8GB VRAM — an RTX 4070 can generate a 1MP image in under 9 seconds via ComfyUI.
Black Forest Labs worked with NVIDIA on quantization paths, which extends to an NVFP4 build for the latest hardware: under 0.6 seconds per 1MP image on a GB300. The model ships in two variants — a 16-step distilled version for fast inference (CFG 1) and a 50-step base for fine-tuning and LoRA training (CFG 3.5). The base model is what the fine-tuning community will care about, and having it available at 5.6B rather than the full model size makes LoRA training meaningfully cheaper.
The license is the FLUX Kommunity License: free commercial use for individuals and organizations under $5M annual revenue; larger deployments require a self-hosted arrangement or the API. Black Forest Labs applies this structure consistently across their releases — open access for individual developers and smaller studios, commercial terms for production at scale. The API starts at $0.009 per 1K images for the distilled model.
Weights are on Hugging Face with ComfyUI and Diffusers support at launch. Black Forest Labs also mentions an external red team evaluation by Cinder, which reportedly found substantially fewer safety vulnerabilities than other popular open-weights image models — a useful data point as these models become more capable at photo-realistic editing.
The pattern here is worth noting. Klein isn’t a compromised open release of a commercial model. It’s a purpose-built compact model that claims to outperform the larger thing. This is different from the usual “you can have the small open weights, the real model is behind the API” story. Whether the quality claims hold up will become clear quickly — the open-weights image generation community runs thorough comparisons fast, and if Klein’s Elo scores are representative, it will show in practice.