NederlandsBaidu releases large AI model ERNIE 5.5 Flash Base, but not for Europe
The Chinese tech company Baidu has released a large AI model that developers can train further themselves. It is built to be fast and economical, and Baidu says it scores well in tests. The licence does not apply in the EU, the United Kingdom and South Korea.
The Chinese tech company Baidu released ERNIE 5.5 Flash Base on 5 November. It is a large AI language model whose weights Baidu is sharing: the numbers that hold everything the model has learned. Developers can therefore download the model from Hugging Face, a widely used website for AI models, and keep working with it themselves.
A model in its base state
What is special is that Baidu is sharing the model in its base state. A language model is made in two phases. First it reads huge amounts of text and learns to predict which word comes next. Then it is refined into an assistant that answers questions and carries out tasks. Baidu is now releasing the model as it is after the first phase. That phase is by far the most expensive. Companies and researchers can do the second phase themselves, for example to train the model for the language of their own field. The version refined as an assistant will follow later, Baidu says.
The model has 280 billion parameters, the adjustable numbers of an AI model. It is built from 512 small sub-networks, each good at something slightly different. For each word, the model switches on only a few of them, together about 14 billion parameters. This keeps the computing work per word much smaller than in a model of the same size that always uses everything. Baidu trained the model on 33.5 trillion pieces of text. After further training, it can process texts of about a million pieces at once, according to Baidu. That is enough for a few thick books.
Economical with memory
Baidu says the model was designed to be economical from the start. While it works, a language model has to remember what came earlier in the text, and with long texts that takes a lot of memory. So only one in seven layers of the model looks at each individual piece of text. The other layers keep a summary for each block of 128 pieces. In addition, the last layer of the model can already guess a few of the next words, so it answers faster.
In Baidu's own tests, the model scores better than comparable base models from DeepSeek, Tencent and Xiaomi on most of the thirty parts. On a well-known knowledge test it reaches 93.3 per cent, against 89.4 per cent for the DeepSeek model. These are figures from Baidu itself; there are no independent measurements yet.
Not for Europe
The model's licence has two notable rules. It does not apply in the European Union, the United Kingdom and South Korea: the model may not be used or distributed there. In addition, companies may use it free of charge as long as their products built with the model earn less than 50 million dollars a year. Above that, they first need written permission from Baidu. The model also may not be used to improve other AI models.
What does this mean for the road to the singularity?
The most expensive part of making a language model is the first training phase. By sharing that step, many more companies and researchers can build a powerful model of their own without spending months on computing. The economical design also shows how large models can be run more cheaply. That Europe is outside the licence for now makes clear that who may use such models also depends on agreements.
What is the technological singularity? Read the explanation →
