FLUX 3 Image Klein: New Capabilities and Specifications

Black Forest Labs has released the FLUX 3 Image Klein model, designed for creating high-quality images with enhanced control and high performance.

The generator architecture includes 5.6 billion parameters, while the Qwen3-VL-4B text encoder is implemented as a separate module. The total model weight volume is approximately 20.3 GB in BF16 format, including the base generator, encoder, and VAE.

The model supports native image generation at resolutions up to 4K and allows using up to ten reference images for content creation or editing.

A key feature is support for bounding boxes, enabling control over object placement, scale, and movement without relying on rigid masks.

Two model versions are available: Base for training and LoRA fine-tuning (50 steps) and Distilled for fast inference (16 steps). Both versions consume similar memory, but the former operates more slowly.

Quantization is supported in BF16, FP8, INT8, and NVFP4 formats. Local deployment is possible on GPUs with at least 8 GB VRAM using an optimized INT8 build in ComfyUI.

Both BF16 versions run stably on the RTX 3090 (24 GB VRAM). Official peak memory usage is 31 GB, though optimized Hugging Face implementations reduce this to 21 GB.

Image generation speed for a 1024x1024 resolution is approximately 4 seconds on the RTX 5090 and 8.9 seconds on the RTX 4070.

The FLUX Kommunity license permits free commercial use of the models for companies with annual revenue under USD 5 million.

Official project resources are available at the following links: