A 10B image model beat an 80B rival. The 'Happy Birthday' test said otherwise.

A 10-billion-parameter open-source image model topped a benchmark that two much larger rivals still lead in production work. The trade-off is the story.

Boogu-Image-0.1 landed on HuggingFace in June 2026 under Apache-2.0, releasing four variants at once: Base for text-to-image, a distilled Turbo for 3–4 step generation, Edit for natural-language editing, and FP8 for quantized deployment. On the Qwen-Image-Bench leaderboard, the 10B Base scored 53.58, edging the 20B Qwen-Image-2512 (52.06) and the 80B Hunyuan-Image-3.0 (50.81) by 1.5 to 2.8 points. Parameter count usually tracks with capability and cost; the inversion is narrow but real.

The official site claims photographic quality and bilingual text rendering. A hands-on test by Leiphone on 2026-07-31 put those claims to work. Base held up on a Wong Kar-wai-style 《花样年华》 poster and a seaside Chinese-New-Year family set, where composition and skin read as photographic. Turbo collapsed on a "Happy Birthday" cake prompt into garbled text. Edit could swap small objects but stumbled on a wide viewpoint change, leaving a person recognizably out of place.

Distillation buys speed at the cost of text fidelity; instruction tuning caps how far a structural edit can travel. The benchmark win and the production limit share one mechanism, not two.

Independent runs on benchmarks beyond Qwen-Image-Bench are still pending, and aggregator coverage frames the release as part of a wider open-source catch-up. A 0.2 release is the next data point.