Infra behind Krea 2: How to train and serve at scale — Gabriel Jorge Menezes, Krea.ai
Krea trained K2 image model from scratch on thousands of GPUs, now open source
“the whole idea was to like kind of bridge the gap between like LLM research and diffusion transformers”
Krea.ai released K2, an image generation model trained entirely from scratch on a large Infiniband-connected GPU cluster, with no base checkpoint. The model is open source with two variants — a raw pre-trained checkpoint for fine-tuning and a turbo post-trained version capable of sub-second inference. The architecture deliberately ports LLM research techniques into diffusion transformers, positioning K2 as a creative exploration tool rather than a polished consumer product.