A 12B model from Paris-based Mistral AI, co-trained with NVIDIA and shipped under Apache 2.0 with a 128k context window and strong coverage of major EU languages. EU-headquartered vendor plus a fully permissive licence make this one of the cleanest open-weight choices for organisations prioritising AI Act and GDPR alignment, with the main caveat being the usual lack of detail on the training corpus.
Sovereignty
Licence: Apache 2.0Commercial: UnrestrictedTraining data: Undisclosed corpusOrigin: EU (France)
Training data sources and composition are not publicly itemised, leaving copyright and personal-data provenance opaque — a recurring weakness for AI Act Article 53 / GDPR DPIA documentation.
Co-trained with NVIDIA on US-operated GPU infrastructure; weights themselves are EU-controlled but the training pipeline touches a non-EU partner.
Downloading from Hugging Face triggers vendor's gated terms and privacy policy, so the acquisition step itself has its own controller relationship to record.
Self-hosting the weights is fully in-scope of the operator's own GDPR controls; using the vendor's hosted endpoint introduces a separate processor relationship that needs a DPA.