Krea AI published a technical report and dropped two downloadable checkpoints — a fine-tunable base and a fast distilled model — under a custom open-weights license.
2026-06-24 · by Moe Ameen
On June 23, 2026, Krea AI published a technical report and released open weights for Krea 2, the first foundation image model the company built from scratch. The release comes in two variants: Krea 2 Raw, an undistilled base checkpoint taken from mid-training and meant for fine-tuning, post-training, and LoRA work, and Krea 2 Turbo, an 8-step distilled model that generates an image in about two seconds. Krea's recommended workflow is to train a LoRA on Raw and run inference on Turbo. Weights are on Hugging Face, code is on GitHub, and the model is reachable through API partners including fal and Replicate, plus community tooling like ComfyUI.
Krea 2 is a text-to-image model tuned for aesthetic range and style control rather than literal prompt-following — a simple prompt yields a wide spread of high-quality looks, and you steer with style references and moodboards. It first launched inside the Krea web product on May 12, 2026, generating in roughly 15 seconds or less on a free tier. By Krea's account and third-party coverage, it ranks near the top of the Artificial Analysis text-to-image leaderboard and is among the strongest image models from an independent lab rather than one of the large foundation labs.
The "open weights" framing carries a caveat worth reading. The custom license is free to use but requires companies above roughly 50 seats to take an enterprise agreement, so it is open weights with a commercial-scale clause rather than fully unrestricted open source. Two other scoping notes: Krea 2 is image-only — Krea the platform offers separate video and 3D tools, but the model itself does not generate video — and at release it targets resolutions up to 1024px, with native 2K/4K, robust editing, and image reference listed as planned rather than shipped.
The practical move this week is to treat Krea 2 as a cheap, fast image source and let Kompozy do everything after the image. Spin up an on-style still in Krea 2 Turbo in a couple of seconds, then drop it into Kompozy as the visual base for a Carousel (brand-exact through HyperFrames), a Quote Graphic, a Photo Post, or a Persona Tweet card. Kompozy writes the captions in your voice via the Persona Brief, generates the formats Krea 2 can't — persona and avatar video, clips, blogs, newsletters — and schedules and publishes the whole set across all nine platforms from one queue. If you run your own keys, Kompozy's BYO-key economics stack neatly with cheap open-weight generation to keep per-post cost low.
There's also a fast content play in the news itself. Open-weight image models are a topic your audience is actively searching this week, so drop your take into Kompozy as a source and let it fan one point of view into a blog post, a carousel explainer, short captioned clips, and platform-native posts, then schedule them across your channels. Being early and clear on a release like this is how a single take becomes a week of content.
Krea 2 is Krea AI's first in-house text-to-image foundation model, tuned for aesthetic range and style control. It launched in the Krea web product on May 12, 2026, and Krea released open weights and a technical report on June 23, 2026.
Raw is an undistilled base checkpoint for fine-tuning and LoRA training. Turbo is an 8-step distilled model that generates in about two seconds. Krea recommends training on Raw and running inference on Turbo.
It is open weights under a custom license — free to use, but companies above roughly 50 seats need an enterprise agreement. That makes it open weights with a commercial-scale clause rather than fully unrestricted open source.
Krea 2 generates the image but does not publish it. Bring the image into Kompozy to build a carousel, quote card, or photo post, write captions in your brand voice, and schedule and publish across nine platforms from one queue.