ERNIE 5.5 uses a mixture-of-experts architecture with 280 billion parameters, of which 14 billion are active for each input. The model was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens. Baidu says it is designed to support up to one million tokens of context during fine-tuning and inference.
Baidu compared the base model with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base across Exam, STEM, Reasoning, Knowledge, Code, Multilingual and Long Context categories, using the same evaluation pipeline. According to Baidu, the model competes with similarly sized open foundation models such as DeepSeek V4 Flash and MiMo V2.5 on reasoning, code and long-context tasks.
Why it matters if you build agents
Developers who want a base model they can tune for their own agents now have a 280-billion-parameter open option with a long context window, though it has no instruction tuning yet.
Because the release has no post-training, agents built on it will need their own fine-tuning or wait for the promised post-trained version before relying on it for tool use or chat behavior.
Baidu's benchmark figures come from its own evaluation, so teams should run their own tests on the tasks their agents perform.
Get stories like this every weekday
One short email with the day's agent tooling news.