Clarity for AI buildersSponsored by AI Routing
availability

Baidu releases ERNIE-5.5-Flash-Base, a 280-billion-parameter model without post-training

Baidu releases ERNIE-5.5-Flash-Base, a 280-billion-parameter model without post-training
Baidu has released ERNIE-5.5-Flash-Base, an open foundation model with 280 billion parameters that ships without fine-tuning or post-training, according to The Decoder. The release leaves developers to adapt the base model for their intended uses; a post-trained version is expected to follow. The model uses a mixture-of-experts architecture, activating 14 billion parameters per input. It was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens. The reported design supports extending the context to as many as one million tokens during fine-tuning and inference—a distinction from the context length used in pre-training. According to Baidu's benchmark results, as reported by The Decoder, ERNIE-5.5-Flash-Base is competitive with similarly sized open foundation models on reasoning, coding and long-context tasks. The reported comparison includes DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, evaluated through the same pipeline across exam, STEM, reasoning, knowledge, code, multilingual and long-context categories. The supplied evidence does not include numerical benchmark scores or independent validation of those claims. The weights are available on Hugging Face in FP8 and BF16 formats, The Decoder reports. The central practical distinction is that this is a base-model release rather than a post-trained assistant: developers receive a foundation for further adaptation, not the promised post-trained version. The supplied report does not specify when that version will arrive or provide licensing terms.