Baidu’s Wenxin 5.5 Flash Preview Released: Wuxia Game Built in C++ in Just 11 Hours

Beating AI Insight News Flash: Baidu releases ERNIE 5.5 Flash Preview. This is the first post-training version of the 5.5 Flash Base, which was open-sourced last week, focusing on enhancing programming, agent, and complex task execution capabilities. The model has a total of 280 billion parameters, with 14 billion activated per inference, and supports a 1 million-token context window. Baidu showcased a 3D wuxia game *Yunxiu Xing* developed by the model. In an 11-hour task, the model used C++ to build a Unreal Engine 5.8 game, and also integrated assets from different engine versions. Notably, ERNIE 5.5 Flash is a pure text model with no access to game visuals; it converted playtest issues—such as character floating, incorrect movement direction, and camera misalignment—into executable code tests, then iteratively revised based on test results. Scenes and characters used pre-built asset packs. Baidu also conducted specialized training for professional software including Blender and CAD. In an official demo, two agents were assigned to develop an airport and an aircraft respectively, then unified the scene, camera, and rendered the animation. The model also created editable W16 engines and lunar rovers in FreeCAD via Python scripts, and can call external tools like Seedance to produce videos. In official released benchmarks, Terminal-Bench 4.0 scored 46.5 points, outperforming comparable domestic models including Hy4 and DeepSeek V4.1 Flash, but still lags behind top-tier models like Claude Opus 5.5. The API is now available on Baidu’s Qianfan platform. During the limited-time 50% discount period, per million tokens cost approximately 1 yuan for input and 4 yuan for output during peak hours, and 0.5 yuan and 2 yuan respectively during off-peak hours. Baidu also pledged to release the Preview model weights under the Apache 2.0 license in the future.