Boogu-Image-0.1 open-sources 10B Apache-2.0 image model family

The interesting thing about the Boogu-Image-0.1 release is not the leaderboard splash, it is the parameter-count math. A 10 billion parameter open-source image family, Apache-2.0 licensed, that reports a Qwen-Image-Bench score of 53.58, ahead of the 20B Qwen-Image-2512 at 52.06 and the 80B Hunyuan-Image-3.0 at 50.81. If that holds up outside the benchmark, the "you need scale" story for open image models needs a rewrite.

The family ships in four variants on GitHub and Hugging Face: Base and Edit at 25 to 50 inference steps, Turbo and Edit-Turbo distilled down to about 4 steps, all with FP8 quantized checkpoints. On the editing side, Boogu's own comparison reports an ImgEdit_O score of 4.64 for the Edit variant, above Nano Banana Pro at 4.37 and Seedream 4.5 at 4.32. On the Boogu Arena eval the authors use, the closed-source ceiling is still visible: GPT Image 2 at 64.69, Nano Banana 2.0 at 59.82, with Boogu positioned "among the very top" of the open field rather than beating the closed leaders.

Why this matters if you are not training image models yourself: the price of a strong open editor just dropped to a download. Teams shipping product photography, poster generation, or bilingual Chinese-English overlays who could not justify the per-image cost of the closed APIs now have an Apache-2.0 alternative that runs on roughly 6 to 24 GB of VRAM depending on the quantization, and a 4-step Turbo path that reportedly hits single inference in under a second on high-end hardware.

The honest caveats. The authors say training data is "roughly one order of magnitude less" than open peers, but the retrieved material does not disclose the absolute image count or the dollar training cost, so take the efficiency framing as their claim, not a settled number. They also acknowledge text rendering is "not stable enough yet" in complex layouts, and an Edit-Turbo hotfix on July 8, 2026 addressed severe image quality degradation and weak removal-task performance, which says the family is still stabilizing.

The direction is the part worth watching. A 10B open model that keeps up with 80B open peers on a public benchmark, and closes on closed-source editors, reframes what a workable in-house image stack has to cost to run.