Baidu releases ERNIE 5.5 Flash Base weights

On November 5, 2026, Baidu’s ERNIE team published ERNIE-5.5-Flash-Base. The release page says it is a sparse mixture-of-experts language model with 280 billion total parameters and 14 billion active parameters per token, pretrained from scratch on 33.5 trillion tokens. Training context extends to 128K tokens. Fine-tuning and inference support a 1 million token context.

The weights are the end-of-pretraining checkpoint. The page says they have not been through mid-training or post-training, and that they are meant as a starting point for domain adaptation, instruction tuning, and reinforcement-learning research. They are on Hugging Face under the ERNIE 5.5 Flash Base Model License Agreement v1.0, in FP8 and BF16 safetensors. The page’s example model name for the BF16 weights is baidu/ERNIE-5.5-Flash-Base-BF16. The FP8 repo linked from the page is baidu/ERNIE-5.5-Flash-Base.

The page says a post-trained ERNIE 5.5 Flash will come afterward. The November 5 post does not give a date for that model. This item is the base-checkpoint release, not a claim that the post-trained model is available.

This directory does not have a published ERNIE or Wenxin tool page, so there is no related-tool card on this item.