Baidu releases ERNIE’s base checkpoint, with geographic and revenue limits

The sparse model has 280 billion total parameters and 14 billion activated per token; Baidu reports pretraining on 33.5 trillion tokens. Its intended value is as a foundation for domain adaptation, instruction tuning and reinforcement-learning research. [1][2][3]

The licence is consequential. Its permitted territory excludes the European Union, United Kingdom and South Korea. The territorial restriction also covers derivative models and use or distribution of their outputs. Within that territory, commercial use with rolling twelve-month related-product or service revenue of US$50 million or less is free of model charges but requires a one-time notice before use and ERNIE attribution. Separate written authorization is required before starting above the threshold or crossing it during use. The definition aggregates entities under common control and counts a related product’s entire revenue even if it uses other models. [3]