ERNIE 5.5 Flash Base · ERNIE

Baidu has released ERNIE-5.5-Flash-Base, a sparse Mixture-of-Experts language model with 280 billion total parameters, 14 billion of which activate per token. The release consists of the raw end-of-pre-training checkpoint, with no mid- or post-training applied, so developers can adapt it for domain-specific work, instruction tuning, or reinforcement-learning research. Baidu designed the architecture and training system together to balance capability against activated compute, KV cache, and long-context costs, emphasizing repository-scale code and long-context understanding in the corpus. The move matters because it gives developers a capable open starting point without bearing the cost of large-scale pre-training. Next, watch how the community fine-tunes the base weights and whether the efficiency claims hold up in downstream training and inference.