The AI Office released preliminary guidelines on April 22, 2025, that change how you determine if your organization qualifies as a General-Purpose AI Model provider under the EU AI Act. If you're developing, fine-tuning, or significantly modifying foundation models, you have until May 22, 2025, to submit feedback through the open consultation. More importantly, assess your compliance status now.
This checklist helps you determine if you're a General-Purpose AI Model provider and understand your obligations under Articles 53 and 55.
What This Checklist Covers
This checklist addresses three critical areas: whether your model qualifies as a General-Purpose AI Model, whether your organization is the provider, and what compliance obligations follow. It applies to organizations developing models from scratch, fine-tuning existing models, or modifying models in ways that might trigger provider status.
The checklist doesn't cover downstream AI system obligations under Articles 16-27. If you're integrating a General-Purpose AI Model into an application or service, you're likely an AI system provider or deployer, not a General-Purpose AI Model provider.
Prerequisites
Before starting this checklist, ensure you have:
- Compute tracking capability: Estimate training compute using either the hardware-based approach (GPU hours, utilization, FLOP/s) or the architecture-based approach (parameters, tokens, FLOP-per-token multiplier). Refer to Annex A.1 of the preliminary guidelines for calculation methods.
- Model lineage documentation: If you're modifying an existing model, maintain records of the original model's compute footprint and your modification compute.
- Release mode clarity: Know whether you're releasing via API, open weights, cloud service, software library, or another distribution channel.
Checklist Items
Model Classification
1. Calculate your model's training compute
Estimate total floating point operations (FLOP) using one of two methods. The hardware-based approach multiplies GPU count, training hours, utilization rate, and theoretical FLOP/s. The architecture-based approach multiplies parameters, training tokens, and FLOP-per-token multiplier (typically 6 for dense transformers).
Good looks like: A documented calculation with source data (GPU logs, parameter counts, token counts) that you can reproduce and defend. If your estimate exceeds 10^22 FLOP for a text or image model, flag it immediately.
2. Determine if your model meets the generality threshold
Assess whether your model displays significant generality and can competently perform a wide range of distinct tasks. Consider whether it can be integrated into various downstream systems or applications.
Good looks like: A written assessment citing the model's capabilities across task categories (e.g., summarization, translation, code generation, question answering). If the model is task-specific or domain-constrained by design, document that limitation.
3. Apply the computational presumption
If your text or image model uses training compute greater than 10^22 FLOP, the AI Office presumes it's a General-Purpose AI Model unless you can demonstrate otherwise.
Good looks like: A clear yes/no determination with supporting compute calculations. If you're below the threshold but believe your model is still general-purpose, document your reasoning.
Provider Status Determination
4. Identify if you're developing a new model or modifying an existing one
Determine whether you're conducting the initial pre-training run or modifying a model that already exists.
Good looks like: A clear lineage statement. For new models: "We are the original provider." For modifications: "We are modifying [original model name] provided by [original provider]."
5. For same-entity modifications: apply the ⅓ threshold
If you're the original provider making subsequent modifications, those modifications create a distinct General-Purpose AI Model if they use more than ⅓ of the presumption threshold (3×10^21 FLOP).
Good looks like: A calculation showing modification compute as a percentage of the threshold. If you're below ⅓, you're updating the existing model. If you're above, you're providing a new model with new compliance obligations.
6. For downstream modifiers: apply the General-Purpose AI Model threshold
If you're modifying someone else's General-Purpose AI Model, you become a provider of a new General-Purpose AI Model if your modifications exceed 3×10^21 FLOP. Your obligations are limited to the modification you conducted (per Recital 109).
Good looks like: A documented calculation of your modification compute. If you exceed the threshold, prepare to comply with Article 53 for your modification. If you're below, you're not a General-Purpose AI Model provider.
7. For downstream modifiers: apply the systemic risk thresholds
If the original model was a General-Purpose AI Model with Systemic Risk, you become a provider if your modification exceeds 3×10^24 FLOP. If the original model was not systemic risk but your cumulative compute (original + modification) exceeds 10^25 FLOP, you become a systemic risk provider.
Good looks like: Two separate calculations: your modification compute alone, and the cumulative compute. If either threshold is met, you assume full Article 55 obligations, not limited to your modification.
Market Placement and Exemptions
8. Determine if you're placing the model on the market
Assess whether you're making the model available through APIs, cloud services, software libraries, or similar distribution channels. Models used solely for internal research, development, or prototyping before market placement are not in scope.
Good looks like: A documented release plan stating the distribution channel and intended users. If you're releasing externally in any form, you're placing it on the market.
9. Evaluate open-source exemption eligibility
Review whether your model meets the definitions for free and open-source licensing, including permissions for access, usage, modification, and distribution. Remember that GPAI models with systemic risk are never exempt, regardless of release method.
Good looks like: A license review confirming that all four permissions (access, usage, modification, distribution) are granted without restriction. If you're systemic risk, document that the exemption does not apply.
Notification and Documentation
10. Notify the Commission of pre-training compute estimates
If you're commencing a large pre-training run, estimate the compute ahead of time and notify the Commission within two weeks of the estimate.
Good looks like: A notification email or submission to the AI Office containing your compute estimate, calculation method, and expected training start date, sent no later than 14 days after you finalize the estimate.
11. Prepare model documentation per Article 53
If you're a General-Purpose AI Model provider, prepare up-to-date Technical Documentation covering training data, model architecture, capabilities, limitations, and copyright policy.
Good looks like: A living document that you can update as the model evolves, structured to address each requirement in Article 53. Include your copyright policy for training data and generated outputs.
12. For systemic risk models: prepare Article 55 compliance
If you're a systemic risk provider, prepare to conduct model evaluations, Adversarial Simulation, serious incident tracking and reporting, and cybersecurity protections.
Good looks like: A compliance plan that maps each Article 55 requirement to an internal process, owner, and timeline. Reference the General-Purpose AI Code of Practice for implementation guidance.
Common Mistakes
Counting inference-time improvements as training compute: The guidelines exclude activities that improve model capabilities at inference time (e.g., retrieval-augmented generation, chain-of-thought prompting). Only count compute that feeds directly into training.
Assuming all fine-tuning makes you a provider: Many fine-tuning runs use far less than 3×10^21 FLOP. Run the calculation before assuming provider status.
Confusing General-Purpose AI Model obligations with AI system obligations: If you integrate a General-Purpose AI Model into an application, you're likely an AI system provider subject to Articles 16-27, not a General-Purpose AI Model provider subject to Article 53. The obligations are different.
Overlooking the cumulative threshold for systemic risk: If you modify a non-systemic model and your combined compute exceeds 10^25 FLOP, you assume full systemic risk obligations. This isn't limited to your modification.
Ignoring the consultation deadline: The May 22, 2025, deadline is your opportunity to influence the final guidelines. If a threshold or definition creates compliance challenges for your organization, submit feedback with specific technical reasoning.
Next Steps
If you've identified that you're a General-Purpose AI Model provider, your immediate actions are:
- Document your compute calculations and model classification determinations
- Review the full preliminary guidelines and Annex A.1 for calculation details
- If you foresee compliance difficulties, initiate dialogue with the AI Office before obligations take effect
- Consider signing the General-Purpose AI Code of Practice to demonstrate adherence and build trust with the Commission
- Submit consultation feedback by May 22, 2025, if any threshold or definition requires clarification for your use case
If you're not a General-Purpose AI Model provider but you're a downstream system provider or deployer, shift your focus to the risk-based obligations in Articles 5, 16-27, and 50. The lines between model and system are drawn precisely for a reason.
