Baidu releases the open-weight ERNIE 5.5 Flash Base checkpoint

Background

ERNIE 5.5 Flash Base is a base model, meaning a model released after pre-training but before the later stages that shape its behavior, such as mid-training and post-training. Developers who receive a base model usually fine-tune it for their own purposes, so a base release is aimed at researchers and builders rather than at end users of a chat application.

Open-weight means the trained parameters are published so that anyone can download and run the model under the terms of its license. This differs from a hosted service, where users only send requests to a provider's servers.

What happened

Baidu introduced ERNIE 5.5 Flash Base on 5 November 2026 and published its weights on Hugging Face.^[1][2] The model is a checkpoint released before mid- and post-training, which means it sits after pre-training and before the final tuning that a deployed model would normally receive.^[2]

Key details

ERNIE 5.5 Flash Base is a 280B-parameter mixture-of-experts model with 14B active parameters, trained on 33.5 trillion tokens.^[2] In a mixture-of-experts design, only part of the network runs for each token, so the 14B active figure describes the computation used per token while 280B describes the total stored parameters.

The model supports up to 1M tokens for fine-tuning and inference, after training at 128K.^[2] A context window is the amount of text a model can consider at one time, measured in tokens.

It is licensed under Baidu's own ERNIE 5.5 Flash Base Model License Agreement v1.0.^[2] Because the license is Baidu's own agreement rather than a standard open-source license, its terms apply specifically to this release and developers should read them before use.

Baidu reported MMLU 93.3 and MMLU-Pro 78.4 for the base model.^[2] MMLU is a widely used benchmark of multiple-choice questions across many academic subjects, and MMLU-Pro is a harder variant of it.

Impact

Releasing the base checkpoint with open weights gives researchers and developers the pre-training stage to examine and adapt, rather than only the finished chat model.