Open weights just got another heavy hitter. Krea dropped the weights for Krea 2 — a 12.9B parameter diffusion transformer trained from scratch on a massive dataset of real images. They released two complementary checkpoints: Raw is the base, undistilled version, good for fine-tuning and training LoRAs; Turbo is an 8-step distilled version that spits out a 2K image in about two seconds on consumer-grade cards.
What’s really rare isn’t just the weights themselves — they also released a full technical report covering everything from data cleaning strategies to distributed training architecture. Usually, frontier labs keep that stuff locked up.
Architecture-wise, it’s a single-stream DiT with 28 blocks, width 6144, using grouped-query attention, SwiGLU, and 3D axial RoPE. The license allows free commercial use for individuals and small teams. I’m curious what card they tested the “two seconds for a 2K image” claim on — has anyone replicated that?