Executive Summary

Key Numbers

Sources: the Hugging Face model card and the LICENSE file in the same repository

Two of the four cards hold numbers the model card leads with. The other two hold numbers that appear only once the license file has been opened separately. The second pair sets the conditions.

Baidu put the weights for ERNIE-5.5-Flash-Base on Hugging Face on November 5, 2026. The model holds 280 billion parameters in total, and 14 billion of them switch on to process any one token. The arrangement is called a mixture of experts: each layer carries 512 experts, and a token activates ten of those plus one shared expert that every token uses. There are 43 layers, 42 of which work that way. Pre-training ran on 33.5 trillion tokens with a training context of up to 128K, and the model card says fine-tuning and inference can stretch that to a million tokens.

The Base at the end of the name marks a model that has not yet been tuned to follow instructions. It arrives less as something to talk to than as material for further training toward a particular purpose. That is also where open weights earn their keep for a company. The model runs inside the building, no data goes out to anyone else's server, and the weights can be retrained on data of one's own.

A scorecard came with the release. Baidu reports that its own evaluation put the model first on 24 of 30 benchmarks. The comparisons run against DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base, all base models of the same class with total parameters between 280 and 310 billion. The figure that catches the eye in the model card's table is 78.4 on MMLU-Pro, where DeepSeek-V4-Flash-Base records 69.2 on the same row. The numbers are self-reported, and even so the table shows how closely the Chinese labs are running to one another.

Next to the model card there is one more file, named LICENSE. Its title reads ERNIE 5.5 Flash Base Model License Agreement v1.0, and rather than reach for a ready-made license such as Apache 2.0, Baidu wrote this one for this model. Section 1 defines twelve terms. The eleventh of them is Territory.

"Territory" means all geographical areas worldwide other than the European Union, the United Kingdom and South Korea.

Everywhere in the world except the European Union, the United Kingdom and South Korea. On its own that is a definition and nothing more, but Section 2 grants the license solely within that Territory. The rights to use the materials, modify them, continue training, build derivative models and distribute them to third parties all arise inside that line. The same sentence conditions the grant on the acceptable use policy in Exhibit A, and Section 1.12 makes that exhibit part of the agreement. Section 5.3 then closes the outside once more.

The Licensed Materials, Derivative Models and their Outputs must not be used, reproduced, modified, distributed or displayed outside the Territory.

Five activities are barred: use, reproduction, modification, distribution and display. The bar covers not only the model and its derivatives but the Outputs they produce. Downloading is itself reproduction, so the safer reading is that a copy pulled down in Korea departs from the clause at that moment.

The restriction does not stop at one generation either. Section 5.2 requires that any agreement governing the use or distribution of the materials or a derivative model incorporate the restrictions in Sections 5.1 and 5.3 and Exhibit A as enforceable terms. That weighs more than telling a recipient once and being done with it. It asks for the same territorial limit to be written into the next contract. Someone who builds a derivative model on these weights and releases it carries the three exclusions along, and a model forked again from that one carries them too. For whoever is on the receiving end, a model whose card never mentions ERNIE 5.5 may still have inherited the condition.

The clause divides by where the activity happens rather than by where the company is incorporated. How far it reaches when a Korean company keeps its servers in a region inside the Territory and runs the model only there does not resolve in a single line. The moment a developer in Korea connects to that server, loads the weights and puts an output on a screen, there is something to argue about over where the use and the display took place. The text leaves that question open.

Being inside the Territory does not exhaust the conditions. Section 4 divides again, this time by money. Non-commercial use is free, with neither a notice nor an authorization required (Section 4.1). Commercial use carries no license charge as long as the annual revenue of the related products or services stays at or below $50 million, though the operator has to send Baidu one notice naming itself, the model version, the products the model goes into, and valid contact information (Section 4.2). The product also has to display "Powered by ERNIE" somewhere. That notice is not an application for permission. Once it goes out, the clause states that qualifying commercial use may proceed without awaiting Licensor's reply, acknowledgment, or approval.

Above $50 million the procedure changes. Section 4.3 requires such a company to obtain separate written commercial authorization before it begins. The clause also covers a company that starts below the line and grows past it: the authorization has to be in hand before the threshold is crossed, and use that lacks it must be suspended. Authorized use keeps the "Powered by ERNIE" notice, and the license does add that use lawfully conducted under Section 4.2 incurs no retroactive charge merely because the threshold is exceeded later. One fact survives all of that. Doing well becomes a reason to renegotiate, and the counterparty is the company that built the model.

Sections 5 and 6 hold several more clauses that hurt to discover late. Section 5.5 forbids using the outputs to improve any other artificial intelligence model or machine-learning system apart from this model and its derivatives, which catches the common practice of generating synthetic data to teach a different model. Section 5.6 requires anyone operating a service that generates outputs to put reasonable technical and organizational safeguards in place and review them periodically. Section 6.6 provides that every license terminates automatically the moment a licensee files a suit against Baidu alleging infringement of intellectual property.

The two gates come in order. Section 2 confines the grant itself to the Territory, so the revenue condition in Section 4 only comes up for whoever is already inside. Outside the Territory it makes no difference whether annual revenue runs above or below $50 million, because the question never arrives. An organization in Korea that opens its review at the revenue threshold has the order backwards.

Korea falling outside an open-weight license is not new this month. In August 2026 the Chinese lab MiniMax released its video generation model H3 and withheld free use of the weights in the United States, the European Union, the United Kingdom and Korea. Developers in those four places have to apply for a license separately, and approval turns on their safeguards and their regulatory standing. Access through the API stayed open everywhere.

MiniMax gave its reasons: regulation that differs country by country and remains unsettled, and the legal fights over copyright in generative AI. The suits Hollywood studios have brought against video generation AI came up in particular. Baidu, for its part, wrote no reason for the three exclusions into the license or the model card.

The two lists are not identical. MiniMax took out four places including the United States, while Baidu left the United States in and took out three. The overlap comes to the European Union, the United Kingdom and South Korea. Two cases make it early to call anything a pattern, and even so it is worth recording that two companies three months apart bundled the same three. Why the European Union and the United Kingdom are tied together is easy enough to guess from the AI Act and the copyright regime there. Why Korea rides along with them does not follow from any public document.

Baidu's own recent history shows the shift most plainly. On June 30, 2025 the company released ten models in the ERNIE 4.5 family under Apache 2.0. The largest of them carried 424 billion total parameters with 47 billion active, which makes it bigger in raw parameter count than this ERNIE 5.5 Flash Base. Apache 2.0 permits commercial use, modification and redistribution without regard to geography. Sixteen months later the same company put a model on the same platform with a contract bearing its own name attached.

The phrase open weights points to one fact only, that the weight files can be downloaded. A separate document lying beside those files decides who may use them, where, and on what conditions. The two travel together without being the same thing, and on a model repository page they show up as a single license line either way. A team that ran ERNIE 4.5 and assumes 5.5 comes on the same terms will have gone wrong not by misreading the scoreboard but by never opening the license field.

Open source is a stricter word. The Open Source Initiative's definition conditions the term on not discriminating against fields of endeavor, persons or groups. A clause that names territories and forbids them fails that condition. Baidu does not use the words open source anywhere in this document either. The title reads Model License Agreement for a reason.

Licenses already get counted when software comes in. Making a list of which packages are present and what conditions ride on each, the work commonly called an SBOM, has moved out of the security team's hands and into the hands of whoever manages the supply chain. For models, many organizations have no such list at all. Weights do not arrive through a package manager the way source code does. They usually arrive because one person downloaded them from a repository and put them on internal storage.

That is where Pebblous reads this as a data supply chain question. Choosing a model begins as a comparison of scores, but from the moment the model comes inside and feeds on company data to become a derivative, it turns into a question of lineage. Without a record of which weights it started from and what conditions were attached to them, nobody can find out later which models a license problem touches. Pulling together the services affected when a single clause changes depends on the same records.

The list of open models running inside the company can be laid out, with five lines to fill in beside each entry. Each line fills in only by opening the license text posted with the repository, and none of it appears on a marketing page.

• Is there a territorial clause? If so, does it turn on where the company is incorporated or on where the activity happens?
• Is there a revenue or user threshold? Below it today, is the procedure for crossing it written down?
• Is there a clause barring the outputs from training other models? Does a plan to generate synthetic data run into it?
• Is there a duty to notify or to display attribution? Who has to be told what, and by when?
• Is there a condition that terminates the license automatically? A patent or intellectual property suit as the trigger is common.

If the answer to even one of the five is that nobody has checked, that place is the blank in the supply chain. Building the list comes before swapping out any one model.

Thank you for reading this far. The clauses themselves can be read in the LICENSE file on Hugging Face, and the model specifications are laid out in the model card. If you have actually built a list of the licenses behind the open models your company runs, we would be glad to hear which entries turned out to be worth having.

Primary Sources

• 1.Baidu. (2026-11-05). "baidu/ERNIE-5.5-Flash-Base." Hugging Face. Model card — source for the specifications and benchmark table.
• 2.Baidu. (2026-11-05). "ERNIE 5.5 Flash Base Model License Agreement v1.0." Hugging Face. License text — source for Section 1.11's Territory clause and the thresholds, restrictions and termination clauses in Sections 4 through 6.

Korean Reporting

• 3.Im, Daejun. (2026-11-07). "바이두, ERNIE 5.5 공개...한국 등 3개국 라이선스서 제외." AI Times.
• 4.AI Times. (2026-08). "미니맥스, 동영상 생성 모델 'H3' 공개...미국·EU·영국·한국은 가중치 자유 사용 제한." The MiniMax H3 precedent — a contrasting case that did state its reasons (regulatory uncertainty, copyright disputes).

International Reporting

• 5.Trending Topics. (2026-11). "Baidu's ERNIE 5.5 Flash Base ... May Not Be Used in the EU." An EU-reader angle, used here for cross-verification.