FLUX.2 Klein 9B Base

Generate and edit images with Black Forest Labs’ undistilled foundation model for maximum flexibility and control.

FLUX.2 Klein 9B Base is a nine billion parameter rectified flow transformer that generates images from text and supports multi-reference editing. This is the full-capacity foundation model without distillation, preserving the complete training signal for fine-tuning, LoRA training, and custom pipelines where control matters more than speed.

What makes this model different

Klein means “small” in German, but this model punches above its weight class. It matches or exceeds models five times its size while running on consumer hardware like the RTX 4090.

This is the base variant, not the distilled version. That means higher output diversity and more flexibility for customization.

How it works

The model uses Black Forest Labs’ FLUX.2 VAE for latent representations, optimized for quality and compression.

Unlike typical image models that separate generation and editing into different checkpoints, Klein 9B Base handles both in a unified architecture. You can generate from text, edit single images, or blend multiple reference images in the same model.

What you can do with it

Text to image generation
Describe what you want and the model generates it. The model follows prompts accurately and handles complex compositions with realistic lighting, proper perspective, and coherent spatial relationships.

Single image editing
Upload an image and describe your changes. The model preserves the structure while applying your edits.

Multi-reference generation
Combine up to 5 reference images to maintain character consistency, product appearance, or visual style across generations. This is useful for creating variations while keeping specific elements consistent.

Training approach

Black Forest Labs filtered pre-training data for not safe for work content and known harmful material. The model went through multiple rounds of safety fine-tuning, with internal and external evaluations focused on preventing abuse.

The base model preserves maximum training signal, making it ideal for research and custom fine-tuning. Higher output diversity than distilled variants means more creative exploration possibilities.

Technical details

• Nine billion parameter rectified flow transformer
• Supports text-to-image, image-to-image, and multi-reference generation
• Runs at higher inference steps than distilled variants
• Released under FLUX Non-Commercial License

When to use this model

Choose Klein 9B Base when you need the full undistilled model for fine-tuning, LoRA training, research projects, or custom pipelines where you want maximum control over the generation process.

If you need faster inference for production applications, the distilled Klein 9B model generates images in under half a second with four inference steps while maintaining quality.

You can try this model on the Replicate Playground at replicate.com/playground