Baidu released ERNIE 5.5 Flash Base, a 280-billion-parameter model, on Hugging Face. The weights are available in FP8 and BF16 formats. The model is shipped without fine-tuning or post-training.
The model uses a mixture-of-experts architecture and was pre-trained on 33.5 trillion tokens. It has 14 billion active parameters per input and supports a 1M-token context window. Baidu claims it competes with DeepSeek V4 Flash and MiMo V2.5.
Developers can download and deploy the model immediately for reasoning, coding, and long-context tasks. Baidu’s decision to ship the weights in raw form gives researchers more control over the training process. This move challenges competitors by offering a powerful open-source option for enterprise applications.
