Baidu offers ERNIE 5.5 Flash Preview through a free OpenRouter option and the Qianfan API, with Apache 2.0 weights planned by November 25. The MoE model activates 14 billion of 280 billion parameters per token and supports up to one million context tokens. The claimed generation rate above 200 tokens per second comes without stated hardware or concurrency conditions. The active-parameter share does not represent total memory requirements, and this intermediate post-training checkpoint and internal evaluations should not be treated as general performance of a finished model.
Back to Tech
Baidu opens free access to ERNIE 5.5 Flash Preview
Online access is distinct from planned weight release, and the 200-token-per-second claim lacks test conditions.