Skip to main content

Ant's InclusionAI Opens Training Books Under EU AI Act

Ant's InclusionAI Opens Training Books Under EU AI Act

Ant Financial's InclusionAI has quietly made a bold move on HuggingFace. The company launched a repository called AI-Transparency, where it's publicly sharing summaries of the training content for its major model families—Ling, Ring, and Ming. This isn't about releasing model weights or datasets. Instead, it answers a simple but thorny question: what data did these models actually learn from?

The repo holds only documents—no weights, no training datasets. But it clearly maps out the origins of models from Ling-2.0 through Ling-3.0, Ring-2.0 to Ring-2.6, plus Ming-Omni and two 16B-A3B multimodal models. Each summary sticks strictly to the model and version it names. Mention Ling-2.0, for instance, and it covers only Ling-2.0 (including Ling1T, Ling-flash2.0, and Ling-mini-2.0). It doesn't spill over to other models in the series or those with quantization suffixes.

A Regulatory Nod, Not a Blog Post

Here's the kicker: this summary format isn't homemade. It follows the "public training content summary template" tied to Article 53(1)(d) of the EU Artificial Intelligence Act (Regulation 2024/1689). In plain terms, this is a standardized disclosure required for providers of general-purpose AI models in Europe. Ant isn't just writing a tech blog—it's checking a compliance box.

The scope is wide. Along the Ling line, you get 2.0, 2.5, 2.6, and 3.0, with 3.0 split into LLM (flash, tiny, flash-fin) and VL (Visual Language) versions like Ling-3.0-flash-VL. Ring goes from 2.0 to 2.5-1T and 2.6-1T. The Ming family includes Omni multimodal models (flash-omni-preview and flash-omni-2.0) plus UniAudio-16B-A3B and UniVision-16B-A3B—two unified audio and visual models. Each summary carries its own version number and update date, describing training content at the specified level. No underlying data is distributed, and no third-party rights are granted.

A word of caution from the officials: these documents aren't an authoritative certification or endorsement. They also don't mean every InclusionAI model is covered. The contact page points to Ant's official developer site for Ling. And the version history reminds readers that dates in the PDFs are document version dates, not initial online release times. For actual changes, check the repository's commit records.

Why This Matters

As global regulation turns "training transparency" from a nice-to-have into a hard threshold, Ant's systematic release of training summaries for its entire model series does two things. First, it paves the way for overseas compliance. Second, it accidentally opens a small window into how their models were developed. For anyone watching AI governance, this is a quiet but significant step.

Key Points:

  • Ant's InclusionAI launched AI-Transparency on HuggingFace, disclosing training summaries for Ling, Ring, and Ming models.
  • The format follows the EU AI Act's Article 53(1)(d) template for public training content summaries.
  • Coverage includes Ling-2.0 to 3.0, Ring-2.0 to 2.6, and Ming-Omni plus two 16B-A3B multimodal models.
  • No weights or datasets are shared; each document is version-specific and doesn't grant third-party rights.
  • Officials note the documents aren't certifications and don't cover all InclusionAI models.