Only 14B Activated Out of 280B Total Parameters: Baidu First Launches the Raw ERNIE 5.5 Flash Base Foundation Model Without Post-Training
Baidu released ERNIE 5.5 Flash Base: a 280B sparse MoE model with 14B activated parameters, trained on 33.5T tokens, launched before any mid-training or post-training processes. It is suitable for academic research and secondary development, but not ideal to be directly used as a consumer chat product.
Baidu has made ERNIE 5.5 Flash Base publicly available. Michal Sutter from MarkTechPost highlighted several key technical figures in his coverage: 280B sparse MoE architecture, 14B activated parameters per token, 33.5T tokens of training data, and no mid-training or post-training applied at the time of public release. Original source link: MarkTechPost.
These metrics are worth breaking down separately. 280B refers to the total sum of all model parameters, while 14B refers to the actual volume of parameters that participate in computation for every individual token. MoE models select different expert modules via a routing system, so it supports a very large total parameter count, but the computational cost for a single inference run is close to that of a standard dense 14B model. Still, VRAM requirements, weight storage, routing logic and parallel engineering design all need to follow the 280B scale system standards. For general readers, you can simply understand it as: the model contains a large number of expert modules, and only a small subset of them are woken up to handle each task.
33.5T tokens is the total scale of the pre-training dataset. This volume means massive investment has been put into the pre-training process, but there are still many steps between a raw foundation model and a production-ready chat product, including instruction fine-tuning, preference alignment, and safety training. The official release material explicitly states the model is published prior to mid-training and post-training stages. If you try to run direct conversations on it, its performance will not match the polished experience of mainstream consumer AI assistants.
For developers, the value of this unprocessed, post-training-free foundation model lies in its huge space for secondary development: you can carry out domain-specific continued pre-training, SFT, routing behavior evaluation, or research on expert utilization patterns of MoE models across different tasks. For regular end users, there is no need to treat it as a ready-to-use conversational chat model at the current stage.
This coverage is sourced from MarkTechPost, authored by Michal Sutter. The original article contains no illustrations, so no supplementary images are added to this post.
发布时间: 2026-11-06 05:33