FLUX.2 [klein] squeezes sub-second image generation onto consumer GPUs

What it does

The repository hosts minimal inference scripts for Black Forest Labs’ FLUX.2 suite. It covers the lightweight FLUX.2 [klein] family—distilled 4B and 9B transformers that generate and edit images in under a second on modern consumer GPUs—and the 32B-parameter FLUX.2 [dev] model for higher-fidelity output. All variants handle text-to-image, single-reference editing, and multi-reference editing through a unified architecture.

The interesting bit

The real maneuver is the licensing and sizing split. The 4B klein model ships under Apache 2.0 and fits inside roughly 8 GB of VRAM, making it a rare fully open-weight, commercially permissive option that competes on latency. Meanwhile, the 32B dev model is gated behind a non-commercial license and essentially requires data-center hardware unless you adopt Hugging Face quantization workarounds.

Key highlights

• FLUX.2 [klein] 4B runs in sub-second inference on hardware like an RTX 3090 or 4070 (~8 GB VRAM).
• All klein and dev models support unified text-to-image, single-reference, and multi-reference editing in one checkpoint.
• The 4B klein models are Apache 2.0; 9B klein, klein Base, and dev use a non-commercial license.
• FLUX.2 [dev] offers prompt upsampling via a local 24B Mistral model or an external OpenRouter API call.
• Built-in optional invisible watermarking and C2PA metadata recommendations for output provenance.

Caveats

• FLUX.2 [dev] requires H100-equivalent VRAM for its standard inference script; consumer GPU support depends on external Hugging Face quantization, not code in this repo.
• The 9B and dev weights are non-commercial, so the Apache 2.0 promise only applies to the 4B variants.
• The repository is explicitly “minimal inference code,” so expect plumbing rather than a polished application framework.

Verdict

Ideal for researchers and integrators who want open weights to hack on, especially the 4B klein model for real-time or edge prototypes. Pass if you need a turnkey commercial product or a unified UI, because this is strictly the engine block.