How to Pick an Open LLM for EU AI Act Compliance

Which open LLMs satisfy the EU AI Act's transparency and provenance requirements for general-purpose AI? A practical decision guide.

1. STEP 1

   Understand the requirements

   EU AI Act Article 53 requires general-purpose AI providers to publish a training-data summary, technical documentation, and respect copyright opt-outs. Article 55 adds systemic-risk obligations for models trained with >10^25 FLOPs.

2. STEP 2

   Check provenance disclosure

   OLMo 2 publishes its full training data — easiest to comply. Mistral Large 3 publishes summaries. Llama 4 and DeepSeek publish summaries but with less detail. Falcon 3 publishes a model card with training data overview.

3. STEP 3

   Verify EU origin / data residency

   Mistral is the only EU-headquartered major open LLM provider; using their model with EU-based inference (Scaleway, OVHcloud) gives the cleanest data residency story. Self-hosting any open model on EU infrastructure satisfies the residency part for most cases.

4. STEP 4

   Document your downstream use

   Even if you self-host an open model, you become a 'deployer' under the AI Act and inherit obligations like risk management, human oversight, and instructions for use. Maintain a Model Use Document for each LLM you deploy.

5. STEP 5

   Prefer Apache-2.0 for the cleanest path

   Apache-2.0 models (Mixtral, Yi, OLMo 2, OpenChat, InternLM) come with no licensor-imposed use restrictions on top of EU obligations. Llama, Gemma, and Qwen layer on additional acceptable-use clauses you must also enforce.