AI startup Krea announced the open‑weight release of its next‑generation image generator, Krea 2, in two checkpoints called "Raw" and "Turbo." Both models are downloadable from Hugging Face under a custom license that requires firms with over 50 seats to pay for enterprise usage and obliges all users to enforce safeguards against illegal, non‑consensual, or defamatory content.
The Raw checkpoint is an undistilled mid‑training version that offers a broad latent space for developers who want deep control and extensive fine‑tuning, though it demands significant compute resources. Turbo, by contrast, is a distilled variant that reduces inference to eight steps, delivering 2K‑resolution images in roughly two seconds on consumer‑grade hardware. Its architecture—built on a 12‑billion‑parameter diffusion transformer with SwiGLU MLPs and grouped‑query attention—maintains high prompt fidelity while cutting latency dramatically. Both models integrate Qwen and FLUX 2 VAEs for fast convergence and include filters to strip synthetic data and harmful material.
Benchmark data from mid‑2026 places Krea 2 Turbo among the quickest open‑weight generators, rivaling proprietary services like Midjourney’s turbo mode while remaining fully customizable. Enterprises can now experiment with the models in‑house, tailoring outputs to brand needs without relying on closed APIs, and Krea plans to expand licensing options as adoption grows.
