Black Forest Labs Releases FLUX 3 Image Klein: 5.6B Open Weights Model for Text-to-Image and Editing

Black Forest Labs has released FLUX 3 Image Klein, a 5.6B parameter open weights model designed for text-to-image generation, image editing, and multi-reference generation (supporting up to 10 reference images). This release shares core capabilities with the full FLUX 3 Image model, including native 4K generation, pixel-perfect editing, and bounding box controls for precise element placement.

Performance and Efficiency

In internal evaluations, FLUX 3 Image Klein ranks first in text-to-image, multi-reference generation, and image editing among open weights models. It also outperforms closed models Ideogram V4.5 and FLUX.2 [pro]. At 5.6B parameters, it is significantly smaller than competitors like Qwen-Image 2.1 (7B) and Krea 2 (12.9B). Compared to the previous FLUX.2 [klein] 9B, the new model gains 91 Elo in text-to-image and 70 Elo in image editing with 38% fewer parameters.

The model is optimized for consumer hardware, requiring 11.2 GB of VRAM in BF16 and 5.6 GB in FP8. With the INT8 build, it runs on NVIDIA GeForce RTX GPUs with as little as 8GB of VRAM. Using the step-distilled model and INT8 build, it generates a 1MP image in approximately 4 seconds on an RTX 5090 and under 9 seconds on an RTX 4070 via ComfyUI.

API and Deployment

Through the BFL API, FLUX 3 Image Klein offers lower latency than Qwen-Image 2.1, FLUX.2 [pro], Ideogram V4.5, and Krea 2, achieving the highest rating at its latency tier. The API pricing starts at $0.009 per 1K image. The model is available under the FLUX Kommunity License, which permits free commercial use for individuals and companies with annual revenue under $5M.

Key deployment features include:

• Weights: Base and distilled models available on Hugging Face.
• Support: Day-one integration with ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion, and vLLM-Omni.
• Training: LoRA training supported via Ostris's AI Toolkit and NVIDIA NeMo AutoModel.
• Safety: Checkpoints were red-teamed internally and evaluated by external partner Cinder, showing fewer vulnerabilities than other popular open weights models.