Skip to main content

Ant's InclusionAI Opens Up About AI Training Data

Ant's InclusionAI Opens Up About AI Training Data

Ant Financial's AI division, InclusionAI, has quietly made a bold move on HuggingFace. They've launched a repository called AI-Transparency, where they're publishing detailed summaries of what data went into training their major model families: Ling, Ring, and Ming. But here's the catch—it's not about sharing weights or datasets. It's about answering one simple question: What exactly did these models learn from?

The Compliance Angle

This isn't just a casual blog post. The summaries follow a strict template tied to Article 53(1)(d) of the EU AI Act—a regulation that forces providers of general-purpose AI models to disclose their training content. In other words, Ant is checking a regulatory box. Each document covers only the specific model version it mentions. For example, if it says Ling-2.0, it only covers Ling-2.0 (including Ling1T, Ling-flash2.0, and Ling-mini-2.0). No blanket statements, no assumptions about other versions.

What's Covered?

The list is extensive. On the Ling side, you've got everything from 2.0 up to 3.0, with 3.0 split into LLM (flash, tiny, flash-fin) and VL (Visual Language) versions. Ring goes from 2.0 to 2.6-1T. And the Ming family includes Omni multimodal models (like flash-omni-preview and flash-omni-2.0) plus two unified audio-visual models: UniAudio-16B-A3B and UniVision-16B-A3B. Each summary comes with its own version number and update date, describing training content at a high level—no raw data, no third-party rights granted.

The Fine Print

Ant is careful to note that these documents aren't certifications or endorsements. They don't cover every InclusionAI model. The contact page points to Ant's official developer site for Ling, and the version history reminds readers that the PDF dates are just document versions—not release dates. For actual changes, you'd need to check the repository's commit records.

Why It Matters

As global regulations turn "training transparency" from a nice-to-have into a hard requirement, Ant's systematic release of training summaries serves two purposes. First, it paves the way for overseas compliance. Second, it offers a rare peek into how their models were built. It's a small window, but a telling one.

Key Points

  • Ant's InclusionAI published training summaries for Ling, Ring, and Ming models on HuggingFace.
  • The move aligns with EU AI Act transparency requirements (Article 53(1)(d)).
  • Documents cover specific model versions only—no weights or raw data shared.
  • The repository includes models from Ling-2.0 to Ling-3.0, Ring-2.0 to Ring-2.6, and Ming Omni plus two 16B-A3B multimodal models.
  • Ant emphasizes these summaries are not certifications and don't cover all models.