🎨 Krea 2 (K2) Model Weights Leak
Weights for the new Krea 2 (K2) image generation model from Krea have appeared online. It is built on a single-stream multimodal diffusion transformer architecture with 12.9 billion parameters and uses Qwen3-VL-4B-Instruct as its text encoder.
🌍 The release of Krea 2 confirms the industry's shift toward multimodal DiT architectures, where visual and textual contexts are integrated into a single stream, significantly improving prompt adherence quality.
👤 Krea can now be run locally via ComfyUI. Thanks to the FP8 version, the model is available on consumer GPUs (16–24 GB VRAM), providing access to high quality without subscriptions.
Source 1: https://x.com/krea_ai/status/2069111451202334885 Source 2: https://huggingface.co/CalamitousFelicitousness/Krea-2-Base-Diffusers