Krea says Krea 2 was its conservative bet
Krea says its $20M GPU cluster will fund Krea 2 and two more in-house models, and calls Krea 2 the safe first step before deliberately riskier training runs.
As we reported earlier, Krea shipped Krea 2 as its first in-house image model. In a blog post answering a Fast Company feature by Mark Wilson, the company has now described what it plans to build next.
Krea says it is running a $20M GPU cluster for a year, and that the spend covers Krea 2 and two more Krea models. Co-founder Victor told Fast Company that Krea 2 is "the conservative model" on that roadmap. The company says it had never trained a model before, and a failed first attempt "would have ended the bet." Krea says the next two models are where it will "be reckless on purpose," trying new training norms and new aesthetics. It did not give timelines or technical details.
The post sets Krea's resources against the largest labs: by its count, $83M raised and 37 staff, compared with about $180B raised by OpenAI and $72B by Anthropic. Krea says it spent seven months building and labeling its own dataset by hand and wrote its own post-training workflows. It argues that post-training, not architecture, is what shapes a model's visual style.
The company also says users can opt their own work out of training and own the IP they generate. It is also experimenting with using AI to credit creators whose style measurably influences an output, which it describes as a possible royalties layer. Krea did not say how that would work or when it might launch.