On November 5, Baidu open-sourced the ERNIE-5.5-Flash-Base pretrained checkpoint: a sparse Mixture-of-Experts architecture with 280 billion total parameters and only 14 billion activated per token. Trained from scratch on 33.5 trillion tokens, it supports a 128K training context and 1M inference context.
Thirty-seven of its 43 layers use compressed attention that caches only one entry per 128 tokens. Both FP8 and BF16 weights are published on Hugging Face. The release adds another heavyweight piece to China's open-source camp, where competition is increasingly measured not just in parameters but in the per-token cost equation that determines real-world deployment economics.
