Baidu releases ERNIE 5.5 Flash Preview
Baidu has released ERNIE 5.5 Flash Preview, a 280B-parameter mixture model with 14B active and a 1M-token context, available now by API and promised as Apache 2.0 open weights.
Baidu announced ERNIE 5.5 Flash Preview, its first post-trained model built on the ERNIE 5.5 Flash Base it released earlier. The model has 280B total parameters, of which 14B are active, plus a 1M-token context window. Baidu says it is strong at coding, 3D modeling, knowledge work and content creation.
The API is live on Baidu's Qianfan platform. The weights will be open-sourced under Apache 2.0 "soon," but Baidu gave no date. The company says the model scores "competitively across major benchmarks" and keeps decoding fast. The announcement material provided here includes no benchmark numbers, pricing or speed figures.
Most of the release is a set of demos. Baidu says the model completed each one end to end inside an agent harness, and it credits the external assets used:
- Unreal Engine game: In a single 11-hour task, the model merged assets from several Unreal Engine versions into one UE 5.8 project and wrote a third-person wuxia game entirely in C++. The model is text-only and cannot see the screen, so it turned playtest bugs into code tests, such as a character hovering above the ground or a misplaced camera.
- Mobile mini-games: Working from design documents, it built nearly 100 games across Cocos, Unity, LayaAir, PixiJS and Canvas 2D. Most use deterministic simulations, so bots can check automatically that every level can be cleared.
- Blender scenes: Two agents used the Blender Python API to build an airport and an airliner. The model then rendered the result with Cycles and OptiX and assembled the video with FFmpeg. A separate demo rebuilt the Hōryū-ji temple complex.
Baidu says it improved game-engine and CAD/Blender skills through domain-specific training data. It also mentions a dedicated evaluation framework for game development but did not release it.