Huawei's Hong Kong AI lab's $400K Boogu model churns out 2K images
Huawei's Hong Kong AI lab open sourced Boogu-Image-0.1, a 10-billion-parameter image generation and editing model. The Boogu team — led by Huawei's Hong Kong AI lab with researchers from several universities — released it with four variants: Base, Turbo, Edit, and Edit-Turbo, supporting text-to-image generation in both Chinese and English.
The base model was trained from scratch on 208.62 million deduplicated images, with a theoretical training cost of roughly $400,000. According to the paper, Boogu lands in the top tier of open-source models on several standard benchmarks, and some results even edge close to leading closed-source systems.
Base and Edit support up to 2K resolution, while Turbo generates 1K images in just three to four steps. The system also interprets and rewrites the user's prompt before routing the task to whichever model fits the difficulty — a small trick that trims unnecessary inference spending.
Model weights, inference code, and some training methods are open under Apache 2.0.