Baidu Bars Its New A.I. Model ERNIE 5.5 From Europe
Set Trending Topics as a preferred source on Google.
All of Europe can download the weights of Baidu’s newest A.I. model; it just isn’t allowed to use them.
The Chinese search giant has just released ERNIE 5.5 Flash Base on Hugging Face, its first large open-weight model since the ERNIE 4.5 family in summer 2025. The license that comes with the model contains a clause rarely seen with open models: the licensed territory covers “all geographical areas worldwide other than the European Union, the United Kingdom and South Korea.” Outside that territory, the model and its outputs “must not be used, reproduced, modified, distributed or displayed.” The acceptable use policy repeats the ban as its very first item.
What the License Actually Prohibits
What matters is where the model is used, not where a company is based. Contrary to some reports, the license does not hinge on where a company has its headquarters or primary operations. It prohibits running, adapting or distributing the model in the EU, the UK and South Korea, or displaying its outputs there. For a European startup, that means in practice: a product for European customers built on ERNIE 5.5 is ruled out, and so is fine-tuning on European servers. Whether an EU company could use the model through a Swiss subsidiary for customers outside Europe is left open by the wording.
Even outside Europe, the license is considerably more restrictive than its predecessor’s, for which Baidu had still chosen the permissive Apache 2.0 license. Commercial products must visibly display “Powered by ERNIE.” Anyone generating up to $50 million in annual revenue with the model must notify Baidu once by email; above that, written authorization is required. And neither the model nor its outputs may be used “to improve any other artificial intelligence model,” a clause aimed at distillation, the practice of training one’s own models on the answers of someone else’s. On top of that come a ban on military use and Hong Kong as the place of jurisdiction.
What the Model Can Do
That makes the restriction all the more striking, because technically ERNIE 5.5 Flash Base is among the leaders. The mixture-of-experts model has 280 billion parameters, about 14 billion of which are active per token, was trained on 33.5 trillion tokens and handles up to one million tokens of context during fine-tuning. In Baidu’s own benchmarks, it beats comparable base models on 24 of 30 tests, for example scoring 93.3 on the MMLU knowledge test versus 89.4 for DeepSeek-V4-Flash-Base, and 82.3 versus 69.5 on the HumanEval+ coding test. There are no independent evaluations yet. It is a pure base model without post-training, intended for research and custom adaptation; a finished chat version is set to follow. Running it requires at least 288 gigabytes of GPU memory.
A direct comparison shows how unusual Baidu’s path is. DeepSeek-V4-Flash, its closest architectural relative with 284 billion parameters, is released under the MIT license, the most permissive there is, and Alibaba’s Qwen models under Apache 2.0. Tencent followed this week with Hy4, a large open model of its own, and Moonshot AI released Kimi K3.1 Pro one day before Baidu. For European developers, the choice of strong Chinese open-weight models therefore remains wide; only ERNIE drops out.
Why Baidu Is Avoiding Europe
Baidu gives no reason. Regulation is the obvious explanation. Since August 2025, the EU AI Act’s obligations for providers of general-purpose models have applied, from technical documentation and a summary of training data to copyright compliance, with risk assessments added for particularly capable models. Exemptions for open models apply only to free licenses, and Baidu has just abandoned those. South Korea’s A.I. framework law took effect in January, and the UK relies on strict data protection rules. Excluding these three markets spares Baidu the obligations and the liability risk.
Baidu is not the first major provider to deliberately leave Europe out. In 2024, Meta barred companies based in the EU from the multimodal Llama 3.2 models, citing the “unpredictable” regulatory environment, and Apple brought Apple Intelligence to Europe months late. Critics of the AI Act see this as proof that the rules are cutting Europe off from the technological frontier. Supporters counter that the obligations are transparency and safety requirements any provider can meet, and that a model whose developer won’t disclose its training data shouldn’t be welcome in Europe anyway. Mistral remains a European alternative, releasing its open models under Apache 2.0.

