Baidu released a large language model (LLM) that outperforms rival models from DeepSeek, Tencent, and Xiaomi in major performance evaluations. In particular, it made public the weights of a model that had undergone only pre-training, enabling developers to build their own AI. However, it excluded South Korea and Europe from the regions where use is permitted.
On the 5th, Baidu released the weights of “ERNIE 5.5 Flash Base (ERNIE 5.5 Flash Base)” through its official blog and on Hugging Face. It is a mixture-of-experts (MoE) model with 280 billion parameters in total, activating only 14 billion when processing tokens to reduce computational burden.
In its own performance evaluation, Baidu said it recorded the highest score in 24 of 30 benchmarks compared with three similarly sized Chinese models, including DeepSeek’s “V4 Flash Base,” Tencent’s “Hy3 Preview Base,” and Xiaomi’s “MiMo V2.5 Base.”
Notably, it scored 78.4 on “MMLU-Pro,” which evaluates complex knowledge and reasoning, ahead of DeepSeek (69.2), Tencent (66.9), and Xiaomi (65.7). It also achieved 82.3 on the coding evaluation “HumanEval+,” higher than the rival models.
However, rival models performed better in Chinese-language knowledge evaluations and some long-document processing tasks. Moreover, these results are from Baidu’s own evaluation, and independent verification by external institutions has not yet been confirmed. Since all the comparison targets are also pre-trained models, the results do not directly demonstrate the performance of finished chatbots used by the general public.
ERNIE 5.5 Flash Base just dropped! Time to open that project folder.
Our efficient sparse MoE language model draws on data spanning multiple domains and languages, achieving SOTA on multiple benchmarks among similarly sized base models. Its architecture and training system… pic.twitter.com/pksev9jR8j
— Baidu Inc. (@Baidu_Inc) November 5, 2026
A notable feature of this release is that it is a “base model (Base Model)” that has not undergone subsequent training such as instruction tuning. Typical conversational AI is tuned after pre-training through instruction tuning, reinforcement learning, and other post-training to tailor its responses to users’ requests. Baidu released the model before such subsequent training so that developers could directly develop AI suited to specific industries or tasks.
This means companies and research institutions can build their own AI models without repeatedly undertaking the extremely costly training of foundation models. Baidu identified specialized-domain training, improving reasoning capabilities, and AI agent development as key areas of application.
The model was pre-trained on a total of 3.35 trillion tokens and supports context lengths of up to one million tokens during additional training and inference. Technology for reducing memory usage was also applied when processing long documents.Baidu also plans to release the fully trained “ERNIE 5.5 Flash” in the future.
This release is interpreted as a move by Baidu to strengthen its foothold in China’s open-weight model competition. Baidu was the first among China’s major tech companies to launch a ChatGPT-like AI chatbot in 2023, but it later failed to attract as much attention as competitors such as DeepSeek and Alibaba gained prominence.
In particular, Baidu founder and CEO Robin Li had long emphasized the competitiveness of closed models, but shifted his strategy toward releasing model weights following DeepSeek’s success last year. This time, Baidu is releasing not only a fully trained model but also a base model for further training, seeking to secure a developer ecosystem.
In China, not only DeepSeek and Alibaba but also Tencent, Xiaomi, and others have recently been actively competing in large-scale model development. Baidu’s direct comparison of the base models of DeepSeek, Tencent, and Xiaomi in this announcement is interpreted as an intention to compete not only in fully trained AI services but also in base model performance and the developer ecosystem.
However, the scope of the model release is subject to considerable restrictions. Baidu’s proprietary license excludes South Korea, the European Union (EU), and the United Kingdom from the regions where use is permitted. Developers in these regions therefore cannot use, modify, or deploy the model under the disclosed license alone. The reason for the exclusions has not been disclosed.
Commercial use is also subject to conditions. If the annual revenue of the related product or service is $50 million or less, it can be used without a separate model usage fee, provided that Baidu is notified in advance and the source is credited. If that amount is exceeded, separate written approval is required. Thus, while this release makes the model weights publicly available, it does not make the model open source in a way that anyone can freely use.
Reporter Im Dae-jun ydj@aitimes.com
