Black Forest Labs logo
Black Forest Labs
Open Weights

FLUX.2 [klein] Base 9B

Released Jan 2026

FLUX.2 [klein] Base 9B is a compact rectified flow transformer model developed by Black Forest Labs, released as part of the FLUX.2 family. As an undistilled foundation model, the 9B Base variant is designed to preserve the complete training signal, offering high output diversity and greater flexibility compared to its step-distilled counterparts. It unifies text-to-image generation and advanced image editing capabilities, including native support for multi-reference workflows where users can provide up to ten images to maintain style or character consistency. \n\nThe model architecture utilizes a 9 billion parameter flow transformer paired with a Qwen3 8B text embedder, enabling a sophisticated understanding of complex, structured prompts and high-fidelity text rendering within images. While the distilled [klein] variants are optimized for sub-second inference, the Base 9B model is intended for high-quality production and research, typically operating on a 25 to 50-step sampling schedule to maximize visual detail and prompt adherence. \n\n## Technical Capabilities and Hardware \nFLUX.2 [klein] Base 9B excels at generating photorealistic textures and physically grounded lighting, specifically tuned to reduce synthetic artifacts. Its unified architecture allows it to perform single-reference and multi-reference editing without requiring separate adapters or specialized checkpoints. Due to its undistilled nature and parameter count, the model typically requires approximately 21GB to 29GB of VRAM, making it accessible on high-end consumer hardware such as the NVIDIA RTX 4090 and professional GPU systems. \n\nThe model is released under the FLUX Non-Commercial License, targeting creative exploration, LoRA training, and academic research. It supports resolutions up to 4 megapixels and includes built-in safety filters for protected content. For optimal results, users are encouraged to provide descriptive, detailed prompts, as the base model does not include automatic prompt enhancement, following input instructions literally to preserve artist intent.

Rankings & Comparison