Krea releases Krea 2 with open weights: a Raw checkpoint, a Turbo

Krea open-sources Krea 2, a 12.9B parameter Diffusion Transformer image model available on Hugging Face in Raw and Turbo checkpoints under a capped license.

Krea is open-sourcing the weights of Krea 2, its first foundation image model trained from scratch, designed for aesthetics, style transfer, and creative control. The model, previously closed on its own platform, is now available as open weights on Hugging Face, in two complementary checkpoints.

Krea 2 Raw is the base model, undistilled and without post-training: diverse and malleable, it is intended for fine-tuning, training LoRAs, and research, but not for direct inference. Krea 2 Turbo is the distilled version with fewer steps, designed to generate quickly and cleanly, up to 2K and on a consumer-grade graphics card. The logic follows an in-house formula: train on Raw, generate on Turbo, with LoRAs learned on the base then applying to Turbo. Both share the same architecture, a 12.9 billion-parameter Diffusion Transformer.

The license, which warrants a closer look, remains. The weights are subject to a Krea 2 Community License which the company describes as permissive, but free commercial use is capped: it is only available to entities generating less than one million dollars in annual revenue, all accounts combined, calculated over a rolling twelve-month period. Beyond that, a separate enterprise license becomes mandatory, to be negotiated with Krea. Additionally, there are substantive conditions: whoever hosts the model must install content, prompt, and output filters; any redistribution must include the license and prefix the name with « Krea »; and the brand prohibits circumventing its watermarking or provenance mechanisms. On the tools side, ComfyUI support and FP8 quantized versions have already appeared.