The FLUX.2 [klein] family of AI models has been released, designed specifically for high-speed, interactive image generation and editing. Promoted as the project's fastest models to date, they are engineered to deliver rapid results while running efficiently on consumer-grade graphics cards.

A key feature highlighted is the unification of text-to-image generation, image editing, and multi-reference composition within a single, compact architecture. The company states this allows for applications requiring real-time iteration, such as interactive design tools or agent-based systems that need visual feedback.

Model Variants and Performance

The release includes several model sizes with different intended uses:

· FLUX.2 [klein] 9B: The flagship "small" model, designed to balance quality and latency, reportedly matching or exceeding the output of models five times its size with sub-second inference.

· FLUX.2 [klein] 4B: A fully open-source model under the Apache 2.0 license, built to run on consumer GPUs like the RTX 3090 with approximately 13GB of VRAM.

· Base Models: Undistilled versions of the 4B and 9B models are also provided, intended for researchers and developers who prioritize fine-tuning and maximum output diversity over raw speed.

The company also released quantized versions (FP8 and NVFP4) developed with NVIDIA, which reduce memory usage and increase inference speed on compatible RTX GPUs.

Licensing and Accessibility

Accessibility is a noted focus. The 4B variant and its base model are released under the permissive Apache 2.0 license for commercial and research use. The more capable 9B models are released under a non-commercial license. All models are available via a production-ready API or for local deployment with open weights.

The announcement positions FLUX.2 [klein] as a step toward "interactive visual intelligence," aiming to enable new applications where speed and responsiveness are as critical as output quality.