Baidu open-sources 280B-parameter ERNIE 5.5 Flash Base
Baidu has released ERNIE 5.5 Flash Base as open weights, bringing a large sparse model and long-context training recipe into China’s open-model race.
What happened
Baidu has released ERNIE 5.5 Flash Base as an open-weight model during the Beijing daytime cycle. The model is reported to contain 280 billion total parameters, with about 14 billion activated per token. Its training used 33.5 trillion tokens, with a 128K training context and support for up to a one-million-token context during fine-tuning and inference.
The release matters because Baidu is making a large, current-generation base model available outside its hosted API stack. Developers can inspect the weights, adapt the model, and evaluate it against other Chinese and international open models. The sparse architecture also makes the headline parameter count less useful as a direct estimate of serving cost: only a fraction of the network is active for each token, but deployment still requires substantial memory and engineering capacity.
Why it matters
Baidu’s move adds another major Chinese model family to the open-weight market at a time when developers are increasingly choosing models based on controllability, local deployment, and fine-tuning flexibility. It also gives researchers a new test of whether a large mixture-of-experts design can translate scale into practical gains without making inference inaccessible.
The important caveat is licensing and reproducibility. Open weights do not automatically mean unrestricted commercial use, and benchmark claims need independent testing across reasoning, coding, multilingual performance, and long-context retrieval. The model’s real influence will depend on whether Chinese cloud platforms, inference vendors, and developer communities make it easy to run.
Baidu is therefore competing on more than model quality: it is trying to turn ERNIE into infrastructure that other builders can depend on.