GPAI Model Technical Documentation: [Model name] [version]

Technical documentation for a general-purpose AI (GPAI) model, organised on Article 53 of the EU Artificial Intelligence Act (Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744, the Digital Omnibus on AI) and its Annex XI (documentation kept for the AI Office and national authorities) and Annex XII (information for downstream providers that integrate the model). It also covers the copyright policy, the public summary of training content and, where the model has systemic risk, the additional Annex XI Section 2 information. Replace every [bracketed] value and every row marked Example. The GPAI provider obligations have applied since 2 August 2025; check the consolidated text for transitional rules and later amendments. This is a documentation format, not legal advice.

Document control

Approvals

Revision history

Contents

1. Scope and regulatory position
2. General description of the model (Annex XI 1)
3. Development process and design (Annex XI 2(a) and 2(b))
4. Training, testing and validation data (Annex XI 2(c))
5. Compute and energy (Annex XI 2(d) and 2(e))
6. Evaluation and adversarial testing
7. Information for downstream providers (Annex XII)
8. Copyright policy and training content summary
9. Systemic risk section
10. Maintenance, requests and contacts

1. Scope and regulatory position

1.1 Model classification

[Explain why the model is a general-purpose AI model: generality and the range of distinct tasks it performs.] Example: A 34-billion-parameter decoder-only language model, trained by self-supervision on text and code, offered through an API and as downloadable weights for many downstream applications.

1.2 Systemic risk

1.3 Obligations matrix

2. General description of the model (Annex XI 1)

2.1 Intended tasks and types of AI systems

2.2 Acceptable use policy

[Link and summary.] Example: The acceptable use policy at [URL], version 3, prohibits use for weapons development, unlawful surveillance, generating child sexual abuse material and deceptive impersonation. API customers accept it in the terms of service; weight licensees accept it in the licence.

2.3 Release date and methods of distribution

2.4 Architecture and number of parameters

2.5 Modality and format of inputs and outputs

2.6 Licence

[Example: API access under commercial terms; weights under the [licence name] licence, version [x], which restricts use per the acceptable use policy. Not released under a free and open-source licence.]

3. Development process and design (Annex XI 2(a) and 2(b))

3.1 Technical means for integration

[Instructions for use, infrastructure and tools needed to integrate the model into AI systems.] Example: REST API with SDKs for Python and TypeScript; self-hosting requires 2 × 80 GB GPUs at 16-bit precision or 1 × 80 GB at 8-bit; reference inference server and quantised weights provided.

3.2 Design specifications and training process

Optimisation objective and what it was meant to achieve: [Example: minimise cross-entropy in pre-training; maximise human preference for helpful and policy-compliant answers in post-training. Relevant parameters: learning rate schedule, batch size, sequence length.]

4. Training, testing and validation data (Annex XI 2(c))

Curation methods: [Example: deduplication (exact and near-duplicate), language identification, quality classifier, removal of personal data patterns, filtering of sources on a blocklist of unsafe sites, removal of evaluation benchmark contamination.]

Detection of unsuitable sources and biases: [Example: source blocklists updated monthly; demographic representation analysis on a 1% sample; toxicity rates per language; documented gaps for low-resource languages.]

Testing and validation data: [Example: held-out sets drawn before training, decontaminated against training data by 13-gram overlap.]

5. Compute and energy (Annex XI 2(d) and 2(e))

For a dense transformer with $N$ parameters trained on $D$ tokens, a common approximation of training compute is

Example: $N = 3.42 \times 10^{10}$, $D = 9.5 \times 10^{12}$ gives $C \approx 1.95 \times 10^{24}$ FLOP for pre-training; post-training and failed runs counted in the total above bring it to $3.1 \times 10^{24}$.

When energy is not metered, estimate it from device-hours $H$, average device power $P$ and data-centre power usage effectiveness:

6. Evaluation and adversarial testing

6.1 Evaluation results

6.2 Adversarial testing

6.3 Known limitations

[Example: hallucination of facts and citations; lower quality in low-resource languages; knowledge cut-off YYYY-MM; vulnerable to prompt injection when processing untrusted content.]

7. Information for downstream providers (Annex XII)

7.1 General description (Annex XII 1)

7.2 Model elements and development (Annex XII 2)

7.3 Integration guidance

[Example: recommended system prompt patterns; safety settings and moderation endpoint; known failure modes; evaluation results relevant to high-risk uses; contact for downstream compliance questions.]

8. Copyright policy and training content summary

8.1 Copyright policy

8.2 Public summary of training content

9. Systemic risk section

[Mark "Not applicable" with the reason if the model has no systemic risk.]

10. Maintenance, requests and contacts

Standard and licence note

Based on Regulation (EU) 2024/1689 (the Artificial Intelligence Act) as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI), Article 53 and Annexes XI and XII, with references to Articles 51, 52 and 55 for models with systemic risk; EU legislative texts are free to reuse. The headings are summarised in original wording and no text of the Regulation or of the General-Purpose AI Code of Practice is reproduced. Check the consolidated text on EUR-Lex and the AI Office for the binding text, delegated acts amending the annexes, templates and guidance. This is a documentation format, not legal advice. Template licence: MIT.