Krea 2: an open image model trained from scratch for variety
Krea released Krea 2, a 12.8B text-to-image model trained from scratch, as a Raw base and an 8-step Turbo. Commercial use is free only below US$1 million in annual revenue.
Krea, the company behind a creative app for images, video and 3D, released Krea 2 on 22 June 2026 and published its technical report the next day. It is a text-to-image model of about 12.8 billion parameters, trained from scratch for what Krea calls creative exploration: a wide range of looks and ways to steer them, not one polished house style. The weights are gated and come with a revenue-capped license.
What is new
The model is a diffusion transformer with grouped-query attention and a Qwen3-VL text encoder. Two checkpoints are open. Krea 2 Raw is the undistilled base: it runs with classifier-free guidance at 52 steps, was trained for images up to about 1K, and is meant for fine-tuning. Krea 2 Turbo is distilled to 8 steps without guidance and draws images from 1K to 2K. Krea recommends training LoRAs on Raw and applying them to Turbo, and Hugging Face already lists more than 1,700 adapters for Turbo.
The report says the pretraining images included no AI-generated ones. The open release makes images from text only. Krea 3, announced on 3 November with editing and up to eight reference images, is hosted, and we found no weights for it.
Running it
Each transformer is 26.3 GB in BF16, and the pipeline adds an 8.9 GB text encoder and a 0.5 GB autoencoder, about 35.7 GB in all. By our estimate that needs about 41 GB of GPU memory without offloading, so this is a workstation-class model, not a laptop one. Diffusers, ComfyUI, SGLang and fal support it. Krea’s prompting guide recommends long natural-language prompts with quotation marks around any text to render, and the repository includes a system prompt for expanding short prompts with an LLM.
License
The Krea 2 Community License is free for commercial use only while your company-wide revenue, counted with affiliates, stays under US$1 million over the last twelve months. Above that you need an enterprise license. Anything you share must start with “Krea” in its name and carry the license and a notice file, you must filter outputs or review them by hand, and the Acceptable Use Policy bans uses such as military work and mass surveillance. Krea can end the license on 30 days’ notice.
Why it matters
Krea 2 sits with FLUX 3 Image Klein in a middle ground: weights you can download and use freely as an individual or a small company, but not permissive. Its US$1 million line is a fifth of FLUX’s US$5 million, so a small business should read the license before it commits.
Krea says the model reached the top 10 of the Artificial Analysis text-to-image leaderboard at launch, and second among independent labs. Those are its own claims, and positions change. We found no independent head-to-head, so see FLUX 3 Klein vs Krea 2 for the differences and test your own prompts.
Sources
- Krea 2 Technical Report (Krea blog, 23 June 2026) krea.ai
- Krea-2-Turbo model card (Hugging Face) huggingface.co
- Krea-2-Raw model card (Hugging Face) huggingface.co
- Krea 2 repository (GitHub) github.com
- Krea 2 license page krea.ai
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Krea 2 Krea | 12.8B | – | Krea 2 Community (Conditional) | 41 GB Workstation | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| Z-Image Tongyi MAI | 6.15B | – | Apache 2.0 (Permissive) | 16 GB Laptop or phone | |
| FIBO 2.0 Bria AI | 8.13B | – | CC BY-NC (Non-commercial) | 29 GB Gaming GPU | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU |