Flux.2 [klein] 9B

All models
Image modelBlack Forest Labs

Overview

Flux.2 [klein] 9B is an efficient 9-billion-parameter image generation and editing model developed by Black Forest Labs. Distilled to require only four inference steps, it delivers high-quality text-to-image outputs and multi-reference edits in under a second. The model is best suited for real-time applications, rapid prototyping, and latency-critical workflows, offering a faster alternative to the provider's larger Flux.2 [pro] and Flux.2 [max] models.

Best of Flux.2 [klein] 9B

What is Flux.2 [klein] 9B best for?

Flux.2 [klein] 9B is Black Forest Labs' flagship compact model, designed for interactive, real-time workflows. It unifies text-to-image generation and multi-reference image editing into a single architecture. The community praises its ability to deliver production-quality photorealism, typography, and character consistency at sub-second speeds. It is highly recommended for rapid prototyping, product marketing assets, and workflows requiring instant visual feedback.

When was Flux.2 [klein] 9B released, and what is its lineage?

Released on January 15, 2026, by Black Forest Labs, this 9-billion parameter distilled rectified flow transformer is the higher-fidelity sibling to the 4B variant. It is part of the broader FLUX.2 family, which succeeded the FLUX.1 generation (such as Flux 1.1 Pro and Flux 1.1 Ultra). The FLUX.2 lineup also includes larger models like Flux.2 [pro], Flux.2 [max], and Flux.2 [flex].

How can I get the best results with Flux.2 [klein] 9B?

Because it is a step-distilled model, you only need 4 inference steps with a Guidance (CFG) scale of 1.0. For prompting, write like a novelist: describe your subject first, followed by the environment, details, and lighting. The model natively supports exact HEX color codes (e.g., #FFB400) and structured JSON prompts for precise layout control. For advanced techniques, consult the official prompting guide.

Similar models

Prompt tips

  • Write Descriptive Prose: Structure prompts like a novelist rather than a search engine. Define the subject first, followed by the environment, style, and technical lighting details.

  • Keep Steps Low: Because the model is step-distilled, it is optimized for exactly 4 inference steps. Pushing the step count higher wastes compute time without improving the output.

  • Leverage Multi-Image Inputs: Use its native multi-reference capabilities by feeding it 2–4 reference images to blend characters, clothing, and styles directly in the prompt.

  • Use KV Caching for Edits: When doing reference-based editing, enable KV caching in your workflow to maintain the logic of the original scene and prevent structural drift.

  • Consider Base Models for Complex Needs: If you need higher output diversity or want to train custom LoRAs, consider stepping up to the undistilled Flux.2 [pro] or Flux.2 [max].