Baidu Wenxin 5.5 Flash Preview Released: Wuxia Game Built with C++ in 11 Hours

Beating AI News Flash: Baidu releases ERNIE 5.5 Flash Preview. This is the first post-training version of the open-sourced 5.5 Flash Base from last week, focusing on enhancing coding, Agent, and complex task execution capabilities. The model has 280 billion total parameters, activates 14 billion parameters per pass, and supports a 1 million Token context.

Baidu demonstrated a 3D wuxia game developed by the model, "Yunxiu Xing." In a single 11-hour task, the model completed Unreal Engine 5.8 game development in C++ and also integrated assets from different engine versions.

Interestingly, ERNIE 5.5 Flash is a pure text model and cannot see the game visuals. It converted issues found in playtest feedback—such as floating characters, incorrect movement directions, and camera misalignment—into executable code tests, then iteratively revised based on test results. The scenes and characters used ready-made asset packs.

Baidu also conducted specialized training for professional software such as Blender and CAD. In the official demo, two Agents divided the work to build an airport and an aircraft, then unified the scene, camera, and rendered the animation. The model also used Python scripts to create an editable W16 engine and lunar rover in FreeCAD, and could call external tools such as Seedance to produce videos.

In the officially published benchmarks, it scored 46.5 on Terminal-Bench 4.0, leading comparable domestic models such as Hy4 and DeepSeek V4.1 Flash, but still lagging behind top models such as Claude Opus 5.5.

The API is now live on Qianfan. During the limited-time 50% off period, per million Tokens, input costs about 1 yuan and output 4 yuan during peak hours, and 0.5 yuan and 2 yuan respectively during off-peak hours.

Baidu also promised to release the Preview weights under the Apache 2.0 license in the future.