Krea 2 — the independent-lab image model with a license asterisk

Krea AI made Krea 2 open-weights public on June 23 — Raw (a CFG-guided base for fine-tuning and LoRA work) and Turbo (an 8-step distilled variant that renders up to ~2K images in a couple of seconds). It is a DiT with a Qwen3-VL text encoder, it runs locally via diffusers with community FP8/GGUF quants already out, and Krea claims the #1 text-to-image slot from an independent lab on Artificial Analysis. The catch is the license: a custom "krea-2-community-license," not Apache/MIT — which is why it lands as a fast take, not a planner pick.

Verdict: An independent lab's first from-scratch image foundation model — locally runnable and very good, but a custom non-OSI license keeps it off the picks

The take

The facts, verified against the Hugging Face cards (`krea/Krea-2-Raw` and `krea/Krea-2-Turbo`): weights were uploaded June 18 and made public June 23, real sharded safetensors, `license: other` → `krea-2-community-license` (a LICENSE.pdf with restrictions for large enterprises — open weights, not OSI-permissive). Architecture is a diffusion transformer with GQA and a Qwen3-VL text encoder; Turbo is the distilled speed variant, Raw is the base meant for downstream fine-tuning. Adoption is real — tens of thousands of downloads in days, and community FP8/NVFP4/GGUF quants already exist, so it is genuinely runnable on a single 24 GB-class GPU.

Why it matters: it is the marquee open-weight image release of the window, and it comes from an independent lab rather than BFL/Alibaba/Stability — a notable datapoint for where open image generation is heading. Quality claims aside, the Raw-plus-Turbo split is the practical story: a base you can fine-tune and a distilled sibling you can actually iterate on at interactive speed.

Our call: no planner-pick change. Our image picks deliberately favor OSI-permissive weights (Z-Image-Turbo Apache, HiDream MIT, FLUX.2 klein 4B Apache), and Krea 2's custom community license — like FLUX.2 dev's non-commercial terms — keeps it a comparator rather than a default recommendation. If you are doing personal or research image work and want a fresh, fast, high-quality option, it is well worth a look; read the LICENSE.pdf before any commercial or large-org use.

Where this fits

Models: FLUX.2 [dev] · Qwen-Image-2512 (20B) + Edit-2511 · HiDream-O1-Image (8B) · Z-Image-Turbo

Hardware: NVIDIA RTX 5090 · NVIDIA RTX 4090 · RTX 5060 Ti 16 GB

Sources

Next step

Try this in the planner→