Krea 2 Just Dropped Open Weights for Both RAW and Turbo — And It's a Huge Deal
The open-source AI image generation landscape just got its biggest shakeup of 2026. Krea AI has officially released the weights for Krea 2 (K2) — not as one model, but as a dual-checkpoint family designed to serve two entirely different workflows.
Two Models, One Vision
K2 RAW is the undistilled base checkpoint. Think of it as the "clay" — diverse, malleable, and built for post-training, fine-tuning, and LoRA adaptation. If you want to customize, experiment, or build your own style from the ground up, RAW is your starting point.
K2 Turbo, by contrast, is the 8-step distilled powerhouse. It generates high-quality images in under 2 seconds, making it one of the fastest open-weight models available today. No waiting, no compromises — just speed and quality running in parallel.
Benchmarks That Turn Heads
Krea 2 is currently ranked #1 independent model on the Artificial Analysis text-to-image leaderboard and sits at #6 globally across all providers. That places it within striking distance of GPT Image 2 and other frontier closed-source offerings — but with the freedom of open weights under a community license.
The team also published a full technical report alongside the release, covering architecture decisions, training methodology, and the distillation process behind Turbo.
Why This Matters
Open-source image models have often trailed proprietary ones by a wide margin — but Krea 2 closes that gap dramatically. With RAW enabling fine-tuning for enterprise use cases and Turbo delivering production-ready speed, this release makes high-quality AI image generation accessible to anyone with a GPU.
Whether you're running inference via ComfyUI, Fal, or building custom LoRAs for a design pipeline — Krea 2 is now the open-source model to beat.
Tags: AI, opensource, machinelearning, design
Links: Krea 2 Open-Source Page | Technical Report | GitHub
Top comments (0)