In a 21-provider census, only DeepSeek and Mistral publish training-data summaries in the EU's template

For 125 days, providers owed a summary "according to a template provided by the AI Office" that the AI Office had not yet provided ^[3][22]. The obligation applied from August 2, 2025 ^[4]. The Commission published the form on December 5, 2025, as C(2025) 8311 final, with an Explanatory Notice attached ^[1].

The form is short, and each field can be checked against a vendor document. Section 1 asks for the provider, the EU authorised representative, a versioned model name and the EU market-placement date ^[1]. It also wants modality, per-modality data-size brackets and types of content ^[1]. Section 2 is a list of data sources, and section 3 covers data processing ^[1]. The Commission calls the template "a common minimal baseline" ^[2].

According to the census, DeepSeek publishes per-model summaries, and its V3.2 summary is essentially that form filled in ^[7]. It is version 1, updated September 3, 2026, and runs five pages ^[7]. It names an EU authorised representative, gives a versioned model ID and per-modality size brackets, and carries a text-and-data-mining reservation ^[7]. Its source list names Common Crawl, Stack Exchange, licensed data and synthetic data, and states no user data ^[7]. The listed EU placement date, August 21, 2025, is 19 days after the obligation started, so this model gets no transition period ^[21]. Versioning the summary per model is the correct engineering choice. A buyer can record which version was reviewed and diff the next one. Mistral's legal center holds lifecycle and technical documents for each of its 42 models, and Large 3 has its own dedicated summary ^[8].

The other 19 providers publish model cards, technical reports or system cards ^[9]. Some are detailed. The census points to Qwen3's token composition, Aleph Alpha's 20T breakdown and Kimi K2's report, yet none of them is a dedicated summary structured against the template ^[9]. Google's Gemini 3 Pro card gives training data one paragraph ^[12]. Meta's Llama 4 card gives it two lines plus token counts ^[13]. Of Anthropic's card, the census authors wrote: "Big safety effort, thin template conformance." ^[15]

The 2-of-21 figure describes the authors' sample. For it to describe a given contract, the vendor has to be in the sample, the model under contract has to be the one checked, and nothing new can have been published since the check. The 21 providers were chosen by the authors' judgment, and they say the sample is non-random ^[16]. Every document was fetched on October 8, 2026 ^[6], 67 days after AI Office enforcement began ^[20]. The top grade requires a dedicated summary whose structure matches the template's three sections ^[17]. The authors concede a provider could plausibly argue that a richer card covers the fields ^[18]. Models placed on the EU market before August 2, 2025 also have until August 2, 2027 ^[5], and the census does not say which of the 19 fall in that window.

In my view the buyer's question is narrow: whether the vendor publishes a document with the template's three sections for the exact model version under contract. On October 8, the public record held DeepSeek's per-model summaries and Mistral's summary for Large 3 ^[6][8].