**Breaking News: Black Forest Labs Unveils FLUX.2 [klein] – Revolutionizing Interactive Visual Intelligence**
Hey everyone, I’ve got some exciting news to share with you all! Black Forest Labs has just dropped FLUX.2 [klein], a game-changing compact picture model that’s part of the FLUX.2 line of interactive visual intelligence models. This new release is all about performance and flexibility, and it’s a major breakthrough for anyone working with AI-powered visual content.
**What’s New with FLUX.2 [klein]?**
If you’re familiar with the FLUX.2 [dev] model, you’ll know it was a 32 billion parameter rectified circulation transformer for text-conditioned image generation and editing. But the new FLUX.2 [klein] takes a similar approach, but with a crucial difference – it’s much more compact. The new model compresses the powerful rectified circulation transformers into smaller 4 billion and 9 billion parameter variants, optimized for rapid inference on modern GPUs.
**Meet the FLUX.2 [klein] Family**
The FLUX.2 [klein] family consists of four principal open-weight variants, each with its own strengths and use cases. The 4B and 9B models are distilled and steering-distilled, making them ideal for production and interactive workloads. The Base variants, with longer sampling schedules, are designed for fine-tuning, LoRA training, and research pipelines.
**Performance and Power**
The FLUX.2 [klein] models offer incredible latency and VRAM performance. The 4B model requires about 13 GB of VRAM, making it suitable for high-end consumer GPUs, while the 9B model requires about 29 GB of VRAM, targeting more powerful GPUs. To take it to the next level, Black Forest Labs has released FP8 and NVFP4 quantized variants, built in collaboration with NVIDIA. These variants can achieve up to 1.6x speedup with 40% fewer VRAM requirements or up to 2.7x speedup with 55% fewer VRAM requirements on RTX GPUs.
**How Does FLUX.2 [klein] Stack Up?**
Black Forest Labs has put FLUX.2 [klein] through its paces with Elo-style comparisons on text-to-image, single-reference editing, and multi-reference generation tasks. The results? FLUX.2 [klein] outperforms QWEN-based models at a fraction of the latency and VRAM, and even surpasses Z Picture in unified text-to-image and multi-reference editing capabilities.
**The Bottom Line**
1. FLUX.2 [klein] is a compact rectified circulation transformer family with 4B and 9B variants supporting text-to-image, single-image editing, and multi-reference generation.
2. The distilled FLUX.2 [klein] 4B and 9B models use four sampling steps and are optimized for sub-second inference on a single modern GPU.
3. Quantized FP8 and NVFP4 variants offer up to 1.6x speedup with 40% fewer VRAM requirements or up to 2.7x speedup with 55% fewer VRAM requirements on RTX GPUs.
For more information, check out the [technical details](https://bfl.ai/blog/flux2-klein-towards-interactive-visual-intelligence), [repo](https://github.com/black-forest-labs/flux2), and [model weights](https://huggingface.co/collections/black-forest-labs/flux2).
