Krea open-sources K2 Raw and K2 Turbo, ranking 2nd among independent image generation labs

Krea released the weights for its Krea 2 image generation models on June 23, 2026, making K2 Raw and K2 Turbo available for download under a permissive license. The release marks Krea's entry into the open-weights image model space, previously dominated by Stability AI's Stable Diffusion family and Black Forest Labs' Flux line.

What's new

The Krea 2 open weights release ships two checkpoints:

• K2 Raw: "An undistilled base model for fine-tuning and post-training research" — intended for researchers and developers who want a clean foundation for post-training, fine-tuning, or style adaptation.
• K2 Turbo: "A guidance- and timestep-distilled model for fast few-step generation" — optimized for production use where speed matters, enabling enterprise-grade image generation at low step counts.

Architecture highlights:

• Diffusion transformer (DiT) design with grouped-query attention (GQA) and sigmoid-gated attention
• SwiGLU MLPs with 4x expansion
• Qwen 3 VL as the text encoder (Alibaba's vision-language model)
• Compatible with both Qwen Image VAE and FLUX 2 VAE autoencoders
• Zero-centered RMSNorm for training stability
• Pretrained across 256px, 512px, and 1024px resolution stages; 2K–4K native generation planned for future releases

Performance: "Krea 2 is among the top 10 models on the Artificial Analysis leaderboard for text-to-image, and scores 2nd place among models from independent labs." The VentureBeat report noted enterprise-grade image generation at approximately 2-second latency.

Access points:

• Weights on Hugging Face (permissive license)
• Code on GitHub
• Official release page at krea.ai/krea-2-open-source
• Interactive demo at krea.ai/image/k2

Context

Krea launched Krea 2 Turbo as a closed product in June 3, 2026, followed by the technical report and open weights release on June 23. The decision to open-weight both the base (Raw) and distilled (Turbo) checkpoints gives the research community access to both ends of the quality-speed tradeoff.

The architecture choice to use Qwen 3 VL as the text encoder is notable — it reflects how Chinese open-weight language models have become default building blocks even for Western AI companies optimizing for multilingual text understanding and image-text alignment. The dual-VAE compatibility (Qwen Image and FLUX 2) also signals interoperability with existing Flux-based tooling in the ComfyUI and A1111 ecosystems.

Why it matters

Open-weights image generation models at this quality tier expand the ecosystem in ways proprietary APIs cannot. Fine-tuning K2 Raw for specific visual styles, product photography aesthetics, or brand-specific outputs is now possible without per-image API costs or data-sharing with a cloud provider. For enterprise teams running sensitive creative workflows, local deployment with K2 Turbo's few-step generation offers a path to 2-second generation speeds on local GPU infrastructure.

The second-place ranking among independent labs — behind only Black Forest Labs' Flux models in that cohort — positions Krea 2 as a serious alternative for developers building creative AI applications who want model quality competitive with Flux but with a different license, architecture, and fine-tuning surface.