Baidu released ERNIE 5.5 Flash Base as an open foundation model. It uses a mixture-of-experts design with 280 billion parameters. Only 14 billion parameters activate per input. The model was pre-trained on 33.5 trillion tokens. It supports a training context of 128,000 tokens. Fine-tuning and inference can extend this to one million tokens. Baidu shipped it without fine-tuning or post-training. Developers must decide how to adapt it themselves. A post-trained version is planned for later. Weights are available on Hugging Face in FP8 and BF16. Benchmarks show it competes with similar open models in reasoning, code and long-context tasks.

• 280 billion parameters
• 14 billion active per input
• 33.5 trillion tokens pre-trained
• 128,000 token training context