inclusionAI Publishes Training Content Summaries for EU Compliance
At a Glance
| Item | Value |
|---|---|
| Repository | inclusionAI/AI-Transparency |
| Published | 2026-09-24 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Overview
On September 24, 2026, inclusionAI released training content summaries for the AI models it provides. This release corresponds to the transparency requirements related to Article 53(1)(d) of Regulation (EU) 2024/1689, a European Union regulation. The target models include a wide-ranging lineup encompassing the Ling, Ring, and Ming series, with documentation provided through the Hugging Face repository “inclusionAI/AI-Transparency".
Announcement Details
The published repository includes PDF-format documents detailing summaries of training content for specific model versions. These do not distribute model weights or the training datasets themselves, but are instead documents intended to explain the composition of the training data and related details.
The main models and versions for which documentation is provided are as follows:
- Ling Series
-
Ling-2.0 (including Ling 1T, Ling-flash 2.0, Ling-mini-2.0) – Ling-2.5 – Ling-2.6-1T (including Ling-2.6-flash and Ling-2.6-1T) – Ling-3.0 LLM (including Ling-3.0-flash, Ling-3.0-tiny, Ling-3.0-flash-fin) – Ling-3.0-flash-VL
-
Ring Series
-
Ring-2.0 (including Ring 1T, Ring-flash 2.0, Ring-mini-2.0) – Ring-2.5-1T – Ring-2.6-1T
-
Ming Series
- Ming-Omni (including Ming-flash-omni-preview and Ming-flash-omni-2.0) – Ming-UniAudio-16B-A3B – Ming-UniVision-16B-A3B
The scope of application for the materials is strictly defined and limited only to the model names and versions identified within each document. If a repository name has a “quantization" suffix, or even if it shares the same model family name, the scope is not automatically extended unless explicitly stated in the document.
Additionally, these summary documents do not constitute the distribution of the underlying training data or grant rights to third-party content. Regarding the specific capabilities, usage methods, limitations, and license terms of the models, it is necessary to refer to the individual repositories for each model. It is explained that the existence of the documents does not imply approval or certification by authorities, nor does it guarantee that all inclusionAI models are covered.
Regarding document management, items are identified by version numbers and update dates within the PDFs. While changes to published files are tracked via repository commit history, the dates written in the PDFs do not necessarily prove the initial publication date online. For any questions, users are guided to inquire through the contact information listed in each summary or the official contact page.
Background
The background behind this documentation release lies in the existence of the Artificial Intelligence Act (Regulation (EU) 2024/1689) adopted by the European Union. This law aims to ensure the safety and transparency of AI systems, and specifically under Article 53(1)(d), obligates providers of general-purpose AI (GPAI) models to make publicly available a sufficiently detailed summary of the content used for training. In compliance with this legal requirement, inclusionAI organized and published the training contents of its major model families such as Ling, Ring, and Ming that it develops and provides.
According to the materials, these summaries were created based on public templates associated with the EU regulation. This is a standard framework designed to enable regulatory authorities and general users to verify what data AI development companies used to build their models. By centrally managing and publishing this information through its “inclusionAI/AI-Transparency" repository, inclusionAI has fulfilled legal compliance while demonstrating a stance of enhancing the transparency of its models. Note that the materials explicitly state that “the existence of the documents does not imply certification by authorities," emphasizing that this is merely a voluntary information disclosure step by the company.
Impact on Local LLM Users
For engineers operating open-weight models in local environments, this announcement provides important information that contributes to the “reliability evaluation" of the models. The fact that the composition of training data, which has tended to be a black box until now, has been articulated as official documentation represents a major step forward.
First, a direct benefit for users is that they can now verify through the published PDF documents what data source composition was used to train models like Ling-3.0 and Ring-2.6 that they download and run. This serves as an important judgment material for understanding the strengths and weaknesses of models, as well as potential information biases included in their outputs. It will likely help when inferring the basis for performance in specific domains from trends in the training data.
At the same time, there are several points that engineers should keep in mind. First, this announcement is merely the publication of “documentation" and does not represent an additional release of model weights or the training datasets themselves. The materials state explicitly that “this repository provides documentation and does not distribute model weights or training datasets." Therefore, the binary resources for building local execution environments do not directly increase.
Second, regarding licenses and availability, the documentation itself neither grants new rights nor imposes restrictions. The materials ask that users continue to refer to the individual repositories of each model for model capabilities, usage methods, limitations, and license terms. Therefore, there is no mention that the terms of use for existing open-weight models are immediately improved or changed by this announcement.
Third is the strictness of the document’s scope of application. According to the materials, the summary scores are limited to the model names and versions identified within each document. If a repository name has a “quantization" suffix, or even if it shares the same model family name, the scope is not automatically extended unless explicitly stated in the document. Users must accurately check whether the specific version they are using is included in the target of the documentation.
Regarding the format of provision, the materials do not reveal policies such as whether there will be a transition to an API-centric approach in the future or whether priority will continue to be given to publishing in formats executable locally. Furthermore, whether future model releases will continue to be open-weight falls outside the scope of this transparency documentation. Overall, this publication serves to fulfill accountability regarding the contents of the models, and it can be said that engineers can now reference more detailed background information when evaluating and selecting models.