Black Forest Labs (BFL) released the weights for FLUX 3 Image Klein on October 15. It is a text-to-image model with 5.6 billion parameters (excluding the text encoder). One model handles text-to-image, image editing and multi-reference generation, accepting up to 10 reference images. Native 4K output, precise editing and bounding-box layout control match what the full FLUX 3 Image offers. It is the second open-weight model in the FLUX 3 family, after FLUX 3 Action.
This site previously covered FLUX 3 Image opening up API access, but that was the full, closed-source version. Klein is the small-size variant that can be downloaded and run locally.
VRAM and speed
BFL is shipping two variants: a base version intended for fine-tuning, and a step-distilled version aimed at low-latency generation. FP8, INT8 and NVFP4 quantized builds are also available. The generator weights come to 11.2GB in BF16 and 5.6GB in FP8, and the INT8 build runs on a GeForce RTX card with as little as 8GB of VRAM.
BFL's local speed figures are quoted per megapixel. With the distilled version at INT8 in ComfyUI, an RTX 5090 takes about 4 seconds per image and an RTX 4070 under 9 seconds. The distilled version produces a 1024×1024 image in 16 steps, ranging from 0.55 seconds on a GB300 to 8.9 seconds on an RTX 4070. The base version runs 50 steps, taking 1.7 seconds on a GB300 and 43.8 seconds on a DGX Spark. Loading a LoRA adds roughly 5% to generation time. Through BFL's own API, median end-to-end latency is 4.54 seconds for text-to-image and 7.18 seconds for image editing.
How it compares with other open models
The open models BFL compares against are all larger than Klein: Qwen-Image 2.1 has roughly 7 billion parameters, Krea 2 has 12.9 billion, and BFL's own previous generation, FLUX.2 [klein], has 9 billion. Against FLUX.2 [klein] 9B, the new model has 38% fewer parameters, a text-to-image Elo score 91 points higher and an image-editing Elo score 70 points higher.
BFL's conclusion is that "FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation and image editing," and that in the same evaluation it also scores above the closed-source Ideogram V4.5 and FLUX.2 [pro]. All of these rankings come from BFL's internal evaluation; third-party blind-test data is not yet available.
In the open-source text-to-image field, Alibaba's Qwen-Image series is Klein's most direct rival, with both sides emphasizing editing and multi-image reference. BFL is betting on size: fewer parameters and a VRAM requirement one tier lower. In the company's words, "Fewer parameters make it practical to run locally and cheaper to adapt." For creators who already run ComfyUI workflows on a single mid-range card, whether it is worth switching will depend on how well it handles Chinese prompts and Chinese text rendering, which BFL did not address in this release.
License and safety
The weights are released under the FLUX Kommunity License. Individuals and companies with annual revenue under $5 million can use the model commercially for free, and research and non-commercial use is also free. Larger organizations need to buy a self-hosting license or go through the API instead. Independent developers and small studios mostly fall within the free tier, and once revenue crosses the threshold a separate licensing deal is required.
On safety, BFL had an outside firm, Cinder, run an assessment before release. The company says Cinder found markedly fewer vulnerabilities than in other common open-weight models, though the assessment method and data have not been made public. The base model was trained on NVIDIA GB200 NVL72 systems.
Sources: Black Forest Labs official blog, CocoLoop; parameter count, VRAM usage, per-hardware generation times and the license revenue threshold were cross-checked.