Bria has trained an 8B-parameter image model using licensed data, matching Qwen in performance; it is open-source but cannot be used directly for commercial purposes.

Beating AI News (Dongcha) – AI image company Bria has launched Fibo 2.0, an 8-billion-parameter model that is the first to integrate text-to-image and image editing capabilities into a single model. Users can edit photos via natural language prompts, or use structured JSON to specify character positions, lighting, and text, minimizing the risk of disrupting other areas when making edits. It also supports mask editing and multi-image reference. Bria has also released a full developer (Dev) version, plus Edit Turbo, an image-editing-accelerated tool that runs inference in just 4 steps. In Artificial Analysis' AA-Image-Editing v2.1 benchmark, Fibo 2.0 tops the open-source model ranking with an Elo score of 1087, slightly edging out Qwen-Image-2.1's 1083. For text-to-image benchmarks, Qwen leads with 1042 points, compared to Fibo's 1039. The gaps in both metrics are not statistically significant, meaning the two models are essentially in the same tier. Fibo 2.0 is trained exclusively on licensed images. Bria has previously partnered with copyright holders including Getty Images, paying for training materials. The model weights are released under the CC BY-NC 4.0 non-commercial license: developers can download them for research and fine-tuning, while commercial use requires a separate agreement with Bria.