Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.
On November 5, 2026, Chinese tech giant Baidu (百度) released the weights (the data that makes up the model) of its AI model "ERNIE-5.5-Flash-Base" on Hugging Face. It is the model as it stands after pre-training (the first stage of training, in which it learns about language from huge amounts of text), and it has not been tuned to answer questions or follow instructions. Baidu says it released the model as a foundation that lets developers start further training for their own fields and goals without paying for large-scale pre-training.
What is a base model?
AI assistants like ChatGPT are built by adding further training after pre-training, such as tuning to follow instructions and reinforcement learning. ERNIE-5.5-Flash-Base is the model at the stage before that. As is, it is not suited to conversation and works by continuing the text it is given.
Baidu cites the following advantages of releasing it in this form.
- Developers are free to choose how to tune it for specific fields, or how to train it for reasoning and agents
- It can serve as a common starting point for research that compares training methods
Baidu says "ERNIE 5.5 Flash," the version tuned to follow instructions, will be released at a later date.
Design and scale
ERNIE-5.5-Flash-Base uses an architecture called "MoE," which runs only the components it needs from a large number of small specialized parts (experts).
- Total parameters: 280B (14B active per token)
- Training data: 33.5T tokens
- Amount of text it can handle at once: up to 128K tokens during training, and 1M tokens for further training and use
- Weight formats: two versions, FP8 and BF16
In 37 of its 43 layers, the model refers to distant text compressed to 1/128 of its size, and only the remaining 6 layers look at the whole text in detail. This keeps computation and memory down when handling long text. Baidu says the training data emphasized code for understanding entire software repositories, as well as long-text comprehension.
Performance
Baidu published results comparing the model under the same conditions with similarly sized base models whose weights are also public: DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base. All are Baidu's own evaluations.
- MMLU-Pro, which tests broad knowledge: 78.4 (DeepSeek 69.2, Tencent 66.9, Xiaomi 65.7)
- LiveCodeBench for programming: 40.1 (31.9, 33.5 and 32.4, respectively)
- GPQA, graduate-level science questions: 60.2 (55.9, 50.2 and 51.4, respectively)
It scored highest of the four models on many evaluations, but models from DeepSeek, Tencent and others came out ahead on some items, including Chinese-language knowledge, long text, and parts of the code and reasoning evaluations.
License and where it can be used
The license is Baidu's own "ERNIE 5.5 Flash Base Model License Agreement v1.0." Its main conditions are as follows.
- Non-commercial use such as research, education and personal experimentation: free
- Commercial use where the related products or services have annual revenue of $50M or less: free, but requires a one-time notice of use to Baidu and a "Powered by ERNIE" label
- Commercial use with annual revenue over $50M: requires written permission from Baidu
- Using the model or its outputs to improve other AI models is prohibited
The license also defines its permitted territory as "worldwide except the EU, the UK and South Korea." Japan is included, but use under this license is not permitted in the EU, the UK or South Korea.
CoAI's take
China's major companies are releasing not only finished AI assistants but also, one after another, the weights of the base models that come before them. The models Baidu compared itself against this time were also base models from DeepSeek, Tencent and Xiaomi, which shows that competition over foundations for further training is intensifying among Chinese players.
On the other hand, a base model cannot be used for chat as is, and further training requires both know-how and computing resources. This release is aimed not at general users but at companies and researchers that want to build their own models. Unlike common open source licenses such as Apache 2.0, it comes with conditions such as territorial restrictions, a revenue threshold and a ban on using it to improve other models, so be sure to read the full license before using it.