Baidu Unveils ERNIE 5.5 Flash Preview 280B Model with Open-Source Pledge

Baidu has released ERNIE 5.5 Flash Preview, the first post-trained version of its Base model series, featuring 280B total parameters with only 14B activated per token, 1M token context window, and generation speed exceeding 200 tokens/s. The API is now available on Baidu's Qianfan platform, and the company has pledged to open-source the weights under Apache 2.0 by November 25.

According to multiple sources, the model employs a sparse architecture optimized for efficient inference, targeting coding, 3D modeling, and content creation tasks that require continuous tool calls. On Arena's Code Arena WebDev leaderboard, it scored 1668, approaching the second spot and slightly outperforming DeepSeek-V4.1-Flash and Kimi K3 while roughly matching Qwen3.8-Max. Baidu claims competitive performance against flagship models like Claude Opus 5.5 and GPT-6.1 Sol on benchmarks such as Terminal-Bench and AutomationBench. However, Baidu acknowledges that as an intermediate post-training version, the model has capability gaps due to insufficient training, with the official release still under rapid iteration.

The release marks a key step for Baidu in efficient inference architecture and open-source strategy. The 280B total/14B active design significantly reduces inference costs, while the 1M context and 200+ tokens/s generation speed target agentic and long-chain tool-calling scenarios. The open-source pledge is expected to attract developer ecosystem and directly compete with DeepSeek, Kimi, and Qwen, intensifying competition in China's top-tier large model landscape.