Europe in 2025: Open, multilingual, and enterprise-ready AI

Europe’s AI scene in 2025 highlights openness, multilingual coverage, and enterprise features driven by research labs and startups across the continent. The models below were selected for technical innovation, practical deployments, and transparent licensing.

Mistral AI (France)

Mistral, founded in Paris in 2023, is notable for open-weight models that emphasize efficiency and scalable architectures like mixture-of-experts (MoE). The company focuses on maximizing performance per parameter and broad context windows for practical applications.

Key models and highlights:

• Mistral Small — ~3.1B parameters, 128k token context, text and image multimodal, licensed Apache 2.0.
• Mixtral (MoE) — 8x7B / 56B MoE, 32k context, strong multilingual performance, Apache 2.0.
• Magistral Small — 1/1.124B, 40k context, reasoning-optimized, Apache 2.0.
• Devstral Small — 124B? entry listed as 124B with 128k context, coding-focused, Apache 2.0.
• Codestral — 12B+, 256k context for advanced code tasks, Apache 2.0.
• Mistral Medium — frontier-tier, 128k context, multimodal and enterprise-focused (API availability).

Strengths: efficient parameter usage, strong coding and reasoning specializations, and open Apache 2.0 licensing for key weights.

Aleph Alpha (Germany)

Aleph Alpha from Heidelberg builds sovereign LLMs with a focus on multilingualism, explainability, and regulatory compliance, aiming to serve public sector and enterprise needs.

Notable models:

• Luminous series — commercial/API offerings tailored to five EU languages with strong semantic representation and embedding features.
• Pharia-1-LLM-7B-Control — 7B, open-source under the Open Aleph License, trained on a multilingual corpus (German, French, Spanish).

Strengths: explainability pipelines, EU AI Act alignment, data sovereignty, and licensing that supports transparent research and non-commercial educational use.

Velvet AI (Italy — Almawave)

Velvet models, developed by Almawave using the Leonardo supercomputer, combine sustainability goals with multilingual coverage and vertical readiness.

Key details:

• Velvet-14B — 14B parameters, 128k context, trained on more than 4T tokens, supports IT/DE/ES/FR/PT/EN, Apache 2.0.
• Velvet-2B — 2B parameters, 32k context, efficient and compact for lighter deployments, Apache 2.0.

Strengths: energy-conscious training choices, broad European language support, and open-source transparency.

Minerva (Italy)

A joint effort by Sapienza NLP, FAIR, and CINECA, Minerva is tailored to Italian language performance while balancing English data.

Model snapshot:

• Minerva 7B — 7.4B parameters, trained on ~2.5T tokens, balanced IT/EN data and instruction-tuned for safer outputs, open-source.

Strengths: strong Italian and English performance with transparent training data and instruction tuning for controlled behavior.

EuroLLM (Pan-European initiative)

EuroLLM aims to provide a truly pan-European foundation by covering official EU languages plus additional regional languages, released in base and instruct variants.

Highlights:

• EuroLLM-9B — 9B parameters, covers 35 languages including all 24 EU official languages, trained on 4T+ tokens, open-source.
• EuroLLM-1.7B — lightweight 1.7B model with the same multilingual coverage, open-source.

Strengths: unmatched open multilingual coverage, competitive translation and reasoning for open-size models, and innovations in dataset balancing like EuroFilter.

LightOn (France)

LightOn provides enterprise-grade, privacy-first solutions with options for on-premises deployment and domain-specific models. The company also explores optical computing approaches.

Representative models and domains:

• Pagnol, RITA, Mambaoutai — general-purpose, open-source offerings.
• Reason-Modern, ColBERT — reasoning and retrieval-focused models.
• BioClinical ModernBERT — biomedical domain models for clinical tasks.

Strengths: private on-prem deployments, domain specialization, and research into efficient hardware-backed inference.

Cross-model comparison and trends

Across these projects, common European priorities include:

• Openness: many models are open-source or provide open weights under Apache or bespoke open licenses.
• Multilingualism: broad language coverage, often including smaller European languages.
• Enterprise readiness: large context windows, reasoning and code-specialized variants, and on-prem privacy options.
• Regulatory and ethical focus: emphasis on explainability, data sovereignty, and compliance with EU rules.

These initiatives together make Europe a distinctive player in 2025, prioritizing inclusive language support, transparent licensing, and enterprise-grade capabilities without sacrificing research-driven innovation.