Baidu has published the weights of ERNIE-5.5-Flash-Base, an open language model with 280 billion total parameters. The release took place on 5 November 2026, and the files can be downloaded from the project's Hugging Face page.

Psst!Add AlexTech.ai asPreferred Source on Google

The model uses a mixture of experts (MoE) architecture, in which not all of the network works on every request: for each token processed, 14 billion parameters are activated, roughly one twentieth of the total. It is a text-only model, designed as a starting point for pre-training and fine-tuning rather than as a ready-to-use assistant.

official image of ERNIE-5.5-Flash-Base

Image: Baidu — https://ernie.baidu.com/release/ernie-5-5-flash-base/

According to the technical documentation, the model was pre-trained on 335 trillion tokens, with a context length of 128,000 tokens at this stage and up to one million tokens after further training. Of its 43 layers, 37 operate on a sequence compressed 128 times, while six dedicated layers preserve performance on long contexts.

Baidu publishes a benchmark table in which the model is very competitive in reasoning, code and exam tests. These are results declared by the company and not yet verified by independent evaluations.

Licence and usage limits

The licence is the point drawing attention. The model is not distributed under a standard open-source licence: the terms of use expressly exclude the European Union, the United Kingdom and South Korea. Those operating in these areas therefore cannot use the weights under the published terms, a restriction that in practice limits the model's diffusion in Europe.

official image of ERNIE-5.5-Flash-Base

Image: Baidu — https://ernie.baidu.com/release/ernie-5-5-flash-base/

This choice sets ERNIE 5.5 Flash Base apart from other open releases in the same family, which mostly carry no geographic restrictions. For Italian developers the model remains available as a technical reference, but cannot be used freely.

The release confirms Baidu's interest in open weights following the ERNIE-Image series, also distributed on Hugging Face, and fits a phase in which more and more Chinese labs publish large models under their own licences.