Baidu Releases ERNIE 5.5 Flash Base with 280 Billion Parameters

Baidu has made available ERNIE-5.5-Flash-Base, an open-source base model whose weights are accessible in FP8 and BF16 on Hugging Face. According to the company, it performs at the same level as other models of comparable size in reasoning, code generation, and handling long contexts. This model, structured as a mixture of experts and totaling 280 billion parameters, with 14 billion activated per request, has benefited from pre-training on 33.5 trillion tokens. A post-trained version is announced, and the model is designed to handle up to one million tokens of context during fine-tuning and inference.

Available Weights and Performance Claims by Baidu

The weights of ERNIE-5.5-Flash-Base are published on Hugging Face in two formats, FP8 and BF16. Evaluations were conducted using the same pipeline and compared against DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base. The tested categories cover Exam, STEM, Reasoning, Knowledge, Code, Multilingual, and Long Context. Baidu states that the model is on par with comparable open-source competitors in reasoning, code, and long context tasks.

Mixture of Experts Architecture and Training Volumes

The model adopts a mixture of experts architecture and totals 280 billion parameters, with 14 billion activated per input. Its pre-training focused on 33.5 trillion tokens. The learning context reached 128,000 tokens.

An Open Base Without Fine-Tuning, with a Post-Trained Version Planned

Distributed as open-source, ERNIE-5.5-Flash-Base is provided without fine-tuning or post-training, allowing developers to determine its direction. The model is designed to support up to one million tokens of context during fine-tuning and inference. Baidu also plans to release a post-trained version.