self-hosted/ai
§01·model · /models

Krea 2

imageactiveKrea 2 Community

Krea AI's first from-scratch, aesthetic-first foundation text-to-image model (released 2026-06-23). ~13B-parameter DiT with flow-matching, paired with a Qwen3-VL text encoder and the Qwen-Image VAE. Two open-weight variants under the Krea 2 Community License: Raw (CFG, full quality; 24.76 GiB BF16 transformer) and Turbo (distilled, ~8-step, ~2 s/image). Runs in ComfyUI 0.25.0+ natively with no custom nodes. The consumer entry point is the FP8 Turbo build (~12 GiB), which fits 16 GB+ GPUs. Commercial use is free for companies under $1M trailing-12-month revenue.

§02·GPUs that run this model
1 total
GPUVRAMSeriesBest speedMin VRAMWorksBenchmarksRecipe
RTX 5060 Ti16GB500recipecheck ↗