# 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.

<!--
Guidance: How to use this template
- Decide first which audience each part serves. Part A (sections 2 to 6)
  follows Annex XI and is provided to the AI Office and national competent
  authorities on request. Part B (section 7) follows Annex XII and is
  given to downstream providers. Mark confidential content clearly.
- The level of detail should match the size and risk profile of the model.
- Open-source models released under a free and open-source licence with
  public weights, architecture and usage information are exempt from the
  Annex XI and XII documentation duties unless they have systemic risk;
  the copyright policy and training content summary still apply. Record
  your position in section 1.3.
- The General-Purpose AI Code of Practice (published by the European
  Commission in July 2025, with Transparency, Copyright, and Safety and
  Security chapters) includes a Model Documentation Form. Providers that
  sign the Code may use it to show compliance; map this file to it.
- Read the consolidated text of the Regulation (including the amendments
  made by Regulation (EU) 2026/1744), its annexes and current AI Office
  guidance for the binding wording; this template summarises the headings
  in its own words.
-->

## Document control

<!-- Guidance: Identify the exact model version, the provider and who is responsible for keeping this documentation current. -->

| Field | Value |
|---|---|
| Document ID | [GPAI-DOC-001] |
| Provider | [Legal name, address, EU authorised representative if established outside the EU] |
| Model name and version | [Name] [version] |
| Model identifiers | [Checkpoint hash, model ID in API] |
| Systemic risk status | [Not systemic risk / Presumed systemic risk (training compute above 10^25 FLOP) / Designated by the Commission] |
| Version of this document | [0.1] |
| Status | [Draft / In review / Approved / Superseded] |
| Owner | [@model-governance-lead] |
| Contributors | [@research-lead], [@data-lead], [@legal], [@safety-lead] |
| Confidentiality | [Part A: confidential, for authorities; Part B: shared under NDA with downstream providers] |
| Last updated | YYYY-MM-DD |

### Approvals

| Role | Name | Decision | Date |
|---|---|---|---|
| Model owner | [@name] | [Approve / Reject] | YYYY-MM-DD |
| Legal / regulatory | [@name] | [Decision] | YYYY-MM-DD |
| Safety lead (systemic risk models) | [@name] | [Decision] | YYYY-MM-DD |

## Revision history

<!-- Guidance: Article 53 requires the documentation to be kept up to date. Each new model version or material change produces a revision. -->

| Version | Date | Author | Change summary |
|---|---|---|---|
| 0.1 | YYYY-MM-DD | [@name] | First draft |
| 1.0 | YYYY-MM-DD | [@name] | *Example:* Documentation for release v2.0 |
| 1.1 | YYYY-MM-DD | [@name] | *Example:* Added evaluation results for fine-tuned v2.1 |

## Contents

<!-- Guidance: Keep aligned with the headings below. -->

1. [Scope and regulatory position](#1-scope-and-regulatory-position)
2. [General description of the model (Annex XI 1)](#2-general-description-of-the-model-annex-xi-1)
3. [Development process and design (Annex XI 2(a) and 2(b))](#3-development-process-and-design-annex-xi-2a-and-2b)
4. [Training, testing and validation data (Annex XI 2(c))](#4-training-testing-and-validation-data-annex-xi-2c)
5. [Compute and energy (Annex XI 2(d) and 2(e))](#5-compute-and-energy-annex-xi-2d-and-2e)
6. [Evaluation and adversarial testing](#6-evaluation-and-adversarial-testing)
7. [Information for downstream providers (Annex XII)](#7-information-for-downstream-providers-annex-xii)
8. [Copyright policy and training content summary](#8-copyright-policy-and-training-content-summary)
9. [Systemic risk section](#9-systemic-risk-section)
10. [Maintenance, requests and contacts](#10-maintenance-requests-and-contacts)

## 1. Scope and regulatory position

<!-- Guidance: Record why the model is a GPAI model, whether it has systemic risk, and which obligations apply. Keep the reasoning so it can be checked. -->

### 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

| Question | Answer | Evidence |
|---|---|---|
| Cumulative training compute | *Example:* 3.1 × 10^24 FLOP | Section 5 |
| Above the presumption threshold in Article 51(2)? | [Yes / No] | |
| Designated by the Commission? | [Yes / No] | |
| Notification to the Commission made (Article 52) | [Yes, YYYY-MM-DD / Not applicable] | |

### 1.3 Obligations matrix

| Obligation (Article 53(1)) | Applies? | Where documented |
|---|---|---|
| (a) Technical documentation (Annex XI) | [Yes / Exempt as open-source without systemic risk] | Sections 2 to 6 |
| (b) Information for downstream providers (Annex XII) | [Yes / Exempt as above] | Section 7 |
| (c) Copyright policy | Yes | Section 8.1 |
| (d) Public summary of training content | Yes | Section 8.2 |
| Article 55 (systemic risk) | [Yes / No] | Section 9 |
| Code of Practice signatory | [Yes: chapters / No: other adequate means] | Section 10 |

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

<!-- Guidance: The six general items Annex XI lists for all GPAI models. Keep them consistent with what is published and what downstream providers receive. -->

### 2.1 Intended tasks and types of AI systems

| Item | Description |
|---|---|
| Intended tasks | *Example:* Text generation, summarisation, question answering, code generation, translation in 24 languages |
| Types of AI systems it can be integrated into | *Example:* Chat assistants, retrieval-augmented search, coding assistants, document processing pipelines |
| Not designed for | *Example:* Autonomous control of physical systems; medical diagnosis without a validated downstream system |

### 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

| Item | Value |
|---|---|
| Release date | YYYY-MM-DD |
| Distribution | *Example:* Public API, cloud marketplaces, downloadable weights on request |
| Regions | *Example:* Worldwide including the EU |

### 2.4 Architecture and number of parameters

| Item | Value |
|---|---|
| Architecture | *Example:* Decoder-only transformer, grouped-query attention, rotary position embeddings |
| Parameters | *Example:* 34.2 billion (dense) |
| Context window | *Example:* 128,000 tokens |
| Tokeniser | *Example:* Byte-level BPE, 128,000 vocabulary |

### 2.5 Modality and format of inputs and outputs

| Direction | Modalities | Formats | Limits |
|---|---|---|---|
| Input | *Example:* Text, images | UTF-8 text; PNG, JPEG | 128k tokens; images up to 8 MP |
| Output | *Example:* Text | UTF-8 text, JSON in structured mode | 8,192 tokens per response |

### 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))

<!-- Guidance: Describe what is needed to integrate the model and how it was designed and trained, including key choices and why they were made. -->

### 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

| Stage | Method | Key choices and rationale |
|---|---|---|
| Pre-training | *Example:* Next-token prediction on 9.5 trillion tokens | Data mix tuned for code and multilingual quality |
| Supervised fine-tuning | *Example:* 1.2 million instruction examples | Mix of licensed, synthetic and human-written data |
| Preference optimisation | *Example:* Direct preference optimisation on 400,000 comparisons | Chosen for stability over reinforcement learning |
| Safety training | *Example:* Refusal and policy data, adversarial examples | Targets acceptable use policy categories |

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))

<!-- Guidance: Describe type and provenance of data, curation methods, volume, scope and main characteristics, how data was obtained and selected, and measures to detect unsuitable sources and identifiable biases. -->

| Dataset category | Type | Provenance | How obtained | Volume | Share of tokens |
|---|---|---|---|---|---|
| *Example:* Public web crawl | Text | Web pages | Own crawler respecting robots.txt and machine-readable opt-outs | 6.1 T tokens | 64% |
| *Example:* Licensed content | Text | News and books publishers | Licence agreements | 0.9 T tokens | 9% |
| *Example:* Public code repositories | Code | Code hosting platforms | Crawler, permissive licences only | 1.6 T tokens | 17% |
| *Example:* Synthetic data | Text, code | Generated by earlier models | Internal pipeline | 0.6 T tokens | 6% |
| *Example:* User data | Text | API customers who opted in | Opt-in | 0.0 (none used) | 0% |
| *Example:* Human annotation | Instructions, preferences | Contracted annotators | Vendor contracts | 1.6 M examples | n/a |

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))

<!-- Guidance: Report the compute used to train the model and the energy consumption, with method. If energy is unknown, estimate it from compute and say so. -->

| Item | Value | Method |
|---|---|---|
| Training compute | *Example:* 3.1 × 10^24 FLOP | Estimated as $6 \cdot N \cdot D$ |
| Training time | *Example:* 41 days | Wall-clock |
| Hardware | *Example:* 4,096 accelerators of type [x] | |
| Energy consumption | *Example:* 2.9 GWh | Metered at cluster level, including cooling |

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

$$
C \approx 6\,N\,D
$$

*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:

$$
E \approx H \cdot P \cdot \text{PUE}
$$

## 6. Evaluation and adversarial testing

<!-- Guidance: Report evaluation results for all models. For models with systemic risk, Annex XI Section 2 asks for a detailed description of evaluation strategies, adversarial testing and, where applicable, system architecture (see section 9). -->

### 6.1 Evaluation results

| Area | Benchmark or method | Metric | Result | Notes |
|---|---|---|---|---|
| Knowledge and reasoning | *Example:* Public multi-subject benchmark | Accuracy | [x] | 5-shot |
| Coding | *Example:* Public code benchmark | pass@1 | [x] | |
| Multilingual | *Example:* Internal 24-language QA set | Accuracy by language | [Range] | Lowest: [language] |
| Safety | *Example:* Internal harmful-request set | Refusal rate on disallowed requests | [x] | |
| Over-refusal | *Example:* Benign borderline set | Unwarranted refusals | [x] | |
| Bias | *Example:* Public bias benchmark | Bias score | [x] | |

### 6.2 Adversarial testing

| Item | Description |
|---|---|
| Teams | *Example:* Internal red team and two external organisations |
| Scope | *Example:* Jailbreaks, harmful instructions, cyber offence uplift, privacy extraction |
| Results summary | [Summary without publishing attack details] |
| Mitigations applied | [List] |

### 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)

<!-- Guidance: What providers who integrate the model need to understand its capabilities and limitations and to meet their own obligations. Annex XII lists general items and development items; protect trade secrets but do not withhold what integrators need. -->

### 7.1 General description (Annex XII 1)

| Item | Value |
|---|---|
| (a) Intended tasks and types of AI systems for integration | See 2.1 |
| (b) Acceptable use policies | See 2.2 |
| (c) Release date and distribution methods | See 2.3 |
| (d) Interaction with hardware or software not part of the model | *Example:* Function calling to external tools defined by the integrator; no built-in internet access |
| (e) Versions of relevant software | *Example:* Reference inference server v1.8; tokenizer library v0.20 |
| (f) Architecture and number of parameters | See 2.4 |
| (g) Input and output modalities and formats | See 2.5 |
| (h) Licence | See 2.6 |

### 7.2 Model elements and development (Annex XII 2)

| Item | Value |
|---|---|
| (a) Technical means for integration | See 3.1; integration guide at [URL] |
| (b) Modalities, formats and maximum size of inputs and outputs | *Example:* Context window 128k tokens; maximum output 8,192 tokens; images up to 8 MP |
| (c) Training, testing and validation data, where applicable | Summary of section 4 types, provenance and curation |

### 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

<!-- Guidance: Article 53(1)(c) requires a policy to comply with Union copyright law, including identifying and honouring rights reservations under Article 4(3) of Directive (EU) 2019/790. Article 53(1)(d) requires a public, sufficiently detailed summary of training content using the AI Office template. -->

### 8.1 Copyright policy

| Element | Description |
|---|---|
| Policy document | [Link, version, owner] |
| Rights reservation detection | *Example:* Crawler honours robots.txt and other machine-readable opt-out protocols; checks before each crawl |
| Lawful access | *Example:* No circumvention of paywalls or technical protection measures |
| Excluded sources | *Example:* Sites listed by authorities as infringing |
| Complaints | *Example:* Rightsholder contact point at [URL]; response within 30 days |
| Output measures | *Example:* Mitigations against reproducing protected content verbatim |

### 8.2 Public summary of training content

| Item | Value |
|---|---|
| Template used | [AI Office template, version] |
| Publication URL | [URL] |
| Published on | YYYY-MM-DD |
| Covers | *Example:* Data categories, main datasets and top domains crawled, data processing aspects |

## 9. Systemic risk section

<!-- Guidance: Complete only if the model has systemic risk. Annex XI Section 2 adds evaluation strategies, adversarial testing and model adaptations, and system architecture where applicable; Article 55 adds evaluation, risk mitigation, serious incident reporting and cybersecurity duties. -->

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

| Item | Description |
|---|---|
| Evaluation strategies and results | *Example:* Protocols, criteria and metrics for dangerous capability evaluations (cyber, chemical and biological, autonomy); methodology for identifying limitations |
| Adversarial testing and model adaptations | *Example:* Internal and external red teaming; alignment and fine-tuning measures adopted as a result |
| System architecture (where applicable) | *Example:* How safety classifiers, the model and serving components connect and feed each other |
| Systemic risk assessment and mitigation | *Example:* Risk register with Union-level systemic risks and mitigations |
| Serious incident tracking and reporting | *Example:* Process to report to the AI Office and national authorities without undue delay |
| Cybersecurity | *Example:* Weight security controls, access logging, insider risk programme |

## 10. Maintenance, requests and contacts

<!-- Guidance: Record how the file is kept current and how authorities and downstream providers get information. -->

| Item | Value |
|---|---|
| Update triggers | New model version; new evaluation results; change of data sources; change of systemic risk status |
| Retention | *Example:* At least 10 years after the model is placed on the market |
| Contact for AI Office requests | [Name, email] |
| Contact for downstream providers | [Name, email, portal] |
| Code of Practice | [Signed chapters, date / not signed: describe alternative means of compliance] |
| Mapping to Model Documentation Form | [Link to mapping table] |

## 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.