FLUX 3 Image Klein: an open image model for 8 GB GPUs

What it does

Klein is the smaller, open version of the full FLUX 3 Image model. BFL says it shares the same core features: native 4K output, pixel-perfect editing and bounding boxes that set where an element goes or confine an edit to one area (BFL).

Two checkpoints are out. The step-distilled model runs 16 steps for fast generation. The base model runs 50 steps and is meant for fine-tuning and LoRA training. Hugging Face says the distilled model produces a 1024×1024 image in about 2 seconds on an H100 (Hugging Face).

Benchmarks

In BFL's internal evaluations, Klein ranks first among open models in text-to-image, multi-reference generation and image editing. The company says it also beats the two closed models in the comparison, Ideogram V4.5 and its own FLUX.2 [pro].

Size is the main argument. At 5.6 billion parameters it is the smallest open model in the comparison, against 7 billion for Qwen-Image 2.1 and 12.9 billion for Krea 2. Compared with FLUX.2 [klein] 9B, it gains 91 Elo in text-to-image and 70 in editing with 38% fewer parameters. Parameter counts cover the image generator only, not the text encoder.

Source: Black Forest Labs. Elo rating versus generator parameters for text-to-image (left) and image editing (right). BFL internal evaluation.

These are the vendor's own measurements. The sources we reviewed do not yet include independent arena results.

What hardware it needs

The generator weights take 11.2 GB in BF16 and 5.6 GB in FP8. The INT8 build, developed with NVIDIA, runs in ComfyUI or Diffusers with 8 GB of VRAM. BFL says the distilled INT8 model generates a 1-megapixel image in about 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070. On a GB300 in NVFP4, an image takes under 0.6 seconds.

Source: Black Forest Labs / NVIDIA. Generation time per GPU for the distilled model at 16 steps. Measurements by NVIDIA.

License and pricing

The weights ship under the FLUX Kommunity License. It allows free commercial use for individuals and companies with annual revenue under $5 million, plus research and non-commercial use for everyone. Larger companies need a self-hosted license or the API.

On the API, the distilled model is available as flux-3-image-klein, starting at $0.009 per 1K-resolution image. Day-one support covers ComfyUI, Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni. Ostris AI Toolkit and NVIDIA NeMo AutoModel provide LoRA training recipes. According to the docs, a LoRA trained on the base model loads into the distilled model without conversion (BFL Docs).

On safety, BFL says the checkpoints were red-teamed internally and evaluated by its external partner Cinder before release.

Who should take a look

Teams that want to generate product shots, ad creatives or mockups on their own hardware without an API bill. Under the $5 million revenue line, you can run and fine-tune it on a single gaming card. Above it, the license conversation comes first. Klein is the second open model in the FLUX 3 family after FLUX 3 Action, and BFL says more will follow.