Baidu unveils ernie 5.5 flash preview model, with weights set for open sourcing by november 25

Baidu has released ERNIE 5.5 Flash Preview, the first fine-tuned iteration based on its previously announced Base model. This new model boasts a total parameter count of 280B and an activation parameter count of 14B, while supporting a 1M context window. The company claims the model performs strongly across various domains, including programming, 3D modeling, knowledge work, and content creation. ERNIE 5.5 Flash Preview is designed to balance high intelligence with efficient inference speed, achieving competitive scores on multiple benchmarks. The model's API is currently available on the Qianfan platform, with the weights scheduled to be open-sourced before November 25. Baidu noted that this preview version is an intermediate step in the post-training process and may still possess some limitations due to incomplete training. The development team is actively continuing the post-training efforts, emphasizing that ERNIE 5.5 is rapidly evolving. They plan to closely monitor the model's real-world performance and user feedback to implement optimizations in subsequent official releases.