Black Forest Labs releases FLUX 3 Image Klein, a 5.6B open-weights model with 4K generation

Black Forest Labs (BFL) has released FLUX 3 Image Klein, an open-weights model with 5.6 billion parameters that, according to the company, handles text-to-image generation, image editing and multi-reference generation — accepting up to ten reference images — at native 4K resolution.

What the model does

BFL positions Klein as sharing its core capabilities with the full FLUX 3 Image model. It can produce 4K output, apply edits while leaving untouched regions of a picture intact, and use bounding boxes either to place new elements at a chosen location or to confine an edit to a defined area.

The model also builds on the wider FLUX 3 family, which BFL says was trained jointly across images, video and audio. Images supply spatial detail, video contributes an understanding of motion and physical dynamics, and audio adds context about events and interactions; combining these signals, the company argues, produces images that better reflect how the world actually looks and behaves.

Two checkpoints ship in the release: a base model intended for fine-tuning, and a step-distilled variant aimed at low-latency generation. BFL also provides FP8, INT8 and NVFP4 builds, developed with support from NVIDIA.

Benchmarks claimed by BFL

The headline numbers come from BFL's own internal evaluations, so they should be read as vendor claims until independently tested. In those evaluations, Klein reportedly ranks first among the compared open-weights models in text-to-image, multi-reference generation and image editing, and also rates above the closed models in the comparison, Ideogram V4.5 and FLUX.2 [pro].

It is also the smallest model in that comparison: 5.6B parameters against 7B for Qwen-Image 2.1 and 12.9B for Krea 2. Compared with its predecessor FLUX.2 [klein] 9B, BFL reports a gain of 91 Elo in text-to-image and 70 in image editing, from 38% fewer parameters.

Through BFL's API, the company additionally claims lower end-to-end latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5 and Krea 2, while achieving the highest quality rating at that latency level.

Built for consumer hardware

The generator weights need 11.2 GB in BF16 and 5.6 GB in FP8, and the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. With the distilled model and the INT8 build, timings measured by NVIDIA show a 1-megapixel image in roughly 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. On an NVIDIA GB300, the distilled model in NVFP4 produces a 1MP image in under 0.6 seconds, and serving a LoRA adds around 5% to generation time.

Ecosystem, license and safety

Weights for both the base and distilled models are available on Hugging Face, alongside reference inference code including a Gradio demo for bounding-box control. Day-one tooling support covers ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training possible through Ostris's AI Toolkit and fine-tuning recipes in NVIDIA's NeMo AutoModel. The distilled model is also served via the BFL API under the name flux-3-image-klein, with features such as prompt upsampling and pixel restoration, and pricing listed from $0.009 per 1K image.

The weights are released under the FLUX Kommunity License: free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use. Larger organisations must either arrange self-hosted licensing or use the API. On safety, BFL says it filtered harmful content out of pre-training data, used targeted fine-tuning after training to suppress harmful behaviours, ran internal red-teaming, and had the checkpoints evaluated by the external safety partner Cinder, which the company says found substantially fewer vulnerabilities than in other popular open-weights models.

Klein is the second open-weights release in the FLUX 3 family, following FLUX 3 Action, a 7B model for robot control, and BFL says further releases are planned.

Why it matters

The gap between open-weights and closed image models keeps narrowing, and Klein's pitch is that it closes while getting smaller rather than larger. A 5.6B model that generates and edits at 4K, runs on an 8 GB consumer GPU and produces a 1MP image in seconds broadens who can deploy capable image AI: hobbyists, small studios and organisations with data-sovereignty or on-device requirements no longer need a data centre or a cloud API. The caveats are real — the benchmark results are BFL's internal evaluations rather than third-party ones, and the revenue-capped licence is not fully permissive for larger companies — but if the numbers hold up outside the vendor's own tests, this release pushes the efficiency frontier of open image generation a meaningful step forward.