FLUX.3 Image Klein lets you control the layout, and its 5.6B weights are open

Black Forest Labs has released FLUX.3 Image Klein, a 5.6B-parameter open-weights model for image generation and editing. The company says its distilled version runs on consumer GPUs with as little as 8 GB of VRAM, and it works with Hugging Face Diffusers from day one.

What is new in FLUX.3 Image Klein

The Hugging Face announcement, written by Aritra Roy Gosthipaty, Sayak Paul and Linoy Tsaban, describes one model that handles text-to-image generation and multi-reference image-to-image editing. It also adds layout control. According to the post, users can specify regions as bounding boxes, written [y_min, x_min, y_max, x_max] on a 0-1000 scale, to place elements or confine edits to a defined area. Black Forest Labs says the model accepts up to 10 reference images and supports native 4K generation.

Two checkpoints are published on the Hugging Face Hub. The default one is time-step distilled and needs 16 steps without classifier-free guidance. The base checkpoint uses 50 steps with guidance and is the one recommended for fine-tuning. Quantized single-file variants of the transformer come in bf16, rowwise FP8 and NVFP4. The Hub lists both checkpoints as 6B, while Black Forest Labs gives 5.6B and counts the text encoders separately.

On the architecture side, the Hugging Face post says the model uses a Qwen3-VL text encoder, the same VAE design as FLUX.2, and a transformer with 28 shared joint blocks. It ships as a Modular Diffusers pipeline with two workflows, text2image and image_conditioned. LoRA training scripts for both text-to-image and editing are included.

Speed, memory and pricing

Hugging Face reports that the distilled checkpoint produces a 1024 x 1024 image in 1.93 seconds on a single H100 80GB, averaged over three runs after warmup. One-reference image-to-image takes 3.87 seconds, and layout editing at 1360 x 768 takes 4.10 seconds. Peak memory after inference is 20.72 GiB for bf16 and 16.27 GiB for FP8 on an H100, and 15.3 GiB for NVFP4 on an RTX 6000. The FP8 kernels are fast on GPUs with compute capability 8.9 or higher, and NVFP4 requires Blackwell hardware.

Black Forest Labs, which built FP8, INT8 and NVFP4 builds with NVIDIA, reports about 4 seconds per megapixel on an RTX 5090 and under 9 seconds on an RTX 4070 using the distilled INT8 model in ComfyUI. On an NVIDIA GB300 with NVFP4, it says generation takes under 0.6 seconds. Generator weights are 11.2 GB in BF16 and 5.6 GB in FP8.

The distilled model is also available through the company’s API as flux-3-image-klein, starting at $0.009 per image, as stated by Black Forest Labs. The API adds prompt upsampling and pixel restoration, which keeps unedited regions unchanged. Weights are on Hugging Face, and a playground and Hugging Face Spaces let users try it. ComfyUI, TensorRT-LLM, SGLang Diffusion and vLLM-Omni also support it from day one.

License and benchmarks

The model is released under the FLUX Kommunity License. According to Black Forest Labs, individuals and companies with annual revenue under $5M can use it commercially for free, and research and non-commercial use are also allowed. Larger organizations need self-hosted licensing or the API.

The company’s own evaluations, not independent tests, put the model first in text-to-image, multi-reference generation and image editing among open-weights models, including larger ones. It also rates it above the closed models Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B, it reports gains of 91 Elo in text-to-image and 70 Elo in editing, with 38% fewer parameters. Comparison models named include Qwen-Image 2.1 and Krea 2. The Hugging Face post itself reports no quality benchmarks.

On safety, Black Forest Labs cites dataset filtering, post-training fine-tuning, internal red teaming and an evaluation by the external partner Cinder, which it says found substantially fewer vulnerabilities than other popular open-weights models.

Why it matters

The first thing to watch is whether the claimed lead holds up outside Black Forest Labs’ own evaluations. The company reports gains over FLUX.2 [klein] 9B, which launched in January 2026. The Decoder reported at the time that the FLUX.2 [klein] 4B model needed about 13 GB of VRAM, so FLUX.3 Image Klein’s 8 GB claim for the distilled version is a real step down. It is still a vendor figure. Independent blind-vote leaderboards and community reports from RTX 4070 and 3090 owners would confirm or contradict it over the coming weeks, especially for layout editing, where bounding-box placement can fail in ways that headline Elo scores do not capture.

The license is the second signal. In the previous generation, The Decoder noted, the 4B model used Apache 2.0, while the 9B model was non-commercial only and businesses needed a separate license. FLUX.3 Image Klein sits between those: it is free for commercial use only below $5M in annual revenue. If community fine-tunes and LoRAs cluster around it, as they did around permissively licensed models, that would suggest the threshold is not a barrier. If developers stay with Apache-licensed alternatives, it would suggest the opposite.

Third, the reported speeds were measured on specific hardware: an H100, an RTX 5090 and a GB300. Real-world numbers on mid-range cards running the INT8 or FP8 builds, and whether the NVFP4 path stays confined to Blackwell GPUs, will show how many people can actually use the model locally.

What’s next

Black Forest Labs says this is the second open-weights release in the FLUX 3 family, after FLUX 3 Action, a 7B robot-control model. It did not announce further releases in the sources reviewed.