A rare combination of permissive licensing, EU provenance, and credible multilingual coverage — Apache 2.0 leaves commercial deployment essentially unconstrained, and Paris-based Mistral simplifies the GDPR controller story compared with US-hosted peers. The catch is the training corpus: Mistral has never published a meaningful dataset breakdown, so AI Act transparency obligations on training-data summaries land squarely on the deployer.
Sovereignty
Licence: Apache 2.0Commercial: UnrestrictedTraining data: UndisclosedOrigin: EU (France)
Training data composition is not disclosed — no dataset card, no opt-out evidence, no copyright provenance statement; AI Act Article 53 training-data summary cannot be reused and must be reconstructed by the deployer.
Sparse MoE inference profile means worst-case memory still tracks the full 141B parameter footprint, which can push deployments onto multi-GPU hosts and complicate on-prem residency claims.
Vendor's `extra_gated_description` on the HF repo signals personal-data processing at download time; review the gate text against your own GDPR notices before pulling weights from a corporate environment.