Physix FrontierPhysix Frontier · News Briefing Card (IT Home · Nov 5, 2026)
Baidu open-sources ERNIE 5.5 Flash Base with 280B total parameters
KEY FACTS
- Baidu's ERNIE team released and open-sourced the ERNIE-5.5-Flash-Base pretraining endpoint weights on November 5.
- The model uses a sparse MoE architecture with 280B total parameters and 14B activated per token.
- The model was pretrained from scratch on 33.5T tokens, with a maximum training context of 128K.
- The fine-tuning and inference stages support a context of up to 1M tokens.
- These weights have not undergone mid-training or post-training; post-trained versions will follow.
- The release gives developers a base checkpoint to build on rather than a finished chat model.
KEY DATA
280BTotal parameters
14BActivated parameters per token
33.5TPretraining tokens
1M tokensMax inference context
PHYSIX OBSERVATION
By open-sourcing base weights that have not been post-trained, Baidu is handing the community the starting point for domain adaptation and reinforcement learning; it is competing for developer ecosystem rather than benchmark rankings. The sparse design of 280B total parameters and 14B activated balances performance and cost, but the real test is whether the post-trained version delivers. For users, there is now one more domestic option for a commercially usable base model.
Source: IT Home report