According to ChainCatcher, Baidu has released the weights for ERNIE-5.5-Flash-Base, a text-only base model with 280 billion total parameters that activates 14 billion per token. Trained from scratch on 33.5 trillion tokens, it supports context windows of up to 1 million tokens. In its 43-layer architecture, 37 layers compress every 128 historical tokens into one cache record, while six retain fine-grained retrieval of the original content to reduce memory and computing costs for long texts. In Baidu’s unified evaluation against DeepSeek-V4-Flash-Base, Tencent Hy3-preview-Base and Xiaomi MiMo-V2.5-Base, the model ranked first in 24 of 30 benchmarks. The release is an untrained base model intended for further fine-tuning, instruction training and reinforcement learning; Baidu said a post-trained version will follow. The company last open-sourced a 100-billion-parameter-class model with the ERNIE 4.5 series in June 2025, about 16 months ago; ERNIE 5.0 and 5.1 were not open-sourced. Unlike ERNIE 4.5’s Apache 2.0 license, ERNIE 5.5 Flash uses a custom license that prohibits use in the EU, UK and South Korea. Products or services that generated more than $50 million in revenue over the past 12 consecutive months require separate written authorization from Baidu for commercial use.