Baidu releases ERNIE 5.5 Flash Preview and promises open weights

Baidu on November 11 released ERNIE 5.5 Flash Preview and opened API access through its Qianfan platform. It is the first post-trained model built on the ERNIE 5.5 Flash Base model the company released earlier this month.

Specifications

The model has 280 billion total parameters, of which 14 billion are active for each token, and supports a one-million-token context window. Baidu describes it as strong at coding, 3D modeling, knowledge work and content creation, balancing capability with fast decoding.

Baidu says the weights will be open-sourced under the Apache 2.0 license; TechNode and other outlets report the release is expected before November 25.

Demonstrations

Baidu published examples the model completed end to end inside an agent harness. In a single 11-hour coding task, it built a third-person game in Unreal Engine 5.8, writing the entire game in C++, and it produced nearly 100 mobile mini-games from design documents. It also wrote FreeCAD scripts to turn architectural references into an editable 3D model and used Blender to block out scenes and camera moves.

Caveats

Baidu notes that the Preview is an intermediate checkpoint from post-training, so some capabilities may still fall short. It plans to incorporate real-world feedback into the official release. The launch continues a trend of major Chinese tech companies publishing capable models with open weights.

Sources: Baidu ERNIE, TechNode