Mistral AI's Large 4 model now sits among 13 European open LLMs with downloadable weights, including Apertus 1.5, EuroLLM-22B, and Bielik 11B v3.1, as of November 12, 2026. The landscape reveals stark differences in licensing, data transparency, and commercial usability.
TL;DR
- 13 European open LLMs now offer downloadable weights, but only a handful provide training data or code.
- Mistral's Large 4 leads with 52B active parameters but restricts commercial API use without a separate license.
- Open models like Apertus 1.5 and EuroLLM-22B balance transparency with partial data/code access.
What happened
As of November 12, 2026, at least 13 European language model families provide downloadable weights, with Mistral AI's Large 4 among the most prominent. Mistral's model features 52 billion active parameters (out of ~1 trillion total) and was released under a modified MIT license on October 29, 2026.
The collection includes models from France (Mistral, Luciole), Switzerland (Apertus), Germany (EuroLLM, Kolibri-1, Teuken), Italy (Velvet2-3B, Domyn Edge, EngGPT 2), Poland (Bielik, PLLuM), and Spain (ALIA-40b). Each model's openness varies: weights are universally available, but training data and code are often restricted.
Mistral's Large 4 requires a separate commercial license for companies offering the model as a service, while internal use remains free. Apertus 1.5, developed by a Swiss public research initiative, provides weights, partial training data, and code under Apache 2.0, but access to larger variants is gated on Hugging Face.
EuroLLM-22B, an EU-funded project, offers Apache 2.0 weights and partial training data but omits the full 4-trillion-token corpus. Bielik 11B v3.1 from Poland's SpeakLeash foundation includes weights and training scripts but requires Hugging Face account consent for access.
Why it matters
For developers, the choice of European open LLMs hinges on licensing and data transparency. Models like Apertus 1.5 and EuroLLM-22B offer more openness but may lack full training data or code, limiting reproducibility.
Startups and enterprises must weigh commercial restrictions—Mistral's Large 4, for example, requires a separate license for API-based services, which could impact adoption costs. The EU's AI Act further complicates compliance by mandating public summaries of training content for some models.
The diversity in licenses and data access reflects Europe's fragmented approach to open AI development. While some models prioritize transparency (e.g., Luciole's GPL-3.0 code), others like EngGPT 2 restrict commercial use entirely, limiting their utility in scalable applications.
Investors should note that the most open models—such as Apertus 1.5—often come with usage policies or indemnity clauses, adding legal complexity. The lack of standardized openness across European LLMs may slow ecosystem growth compared to the more unified approaches seen in the U.S. or China.
Key facts
- As of November 12, 2026, 13 European LLM families offer downloadable weights, including Mistral Large 4, Apertus 1.5, EuroLLM-22B, and Bielik 11B v3.1.
- Mistral Large 4 has 52 billion active parameters (out of ~1 trillion total) and was released on October 29, 2026 under a modified MIT license.
- Apertus 1.5, developed by ETH Zurich, EPFL, and CSCS, provides weights, partial training data, and code under Apache 2.0, but larger variants require Hugging Face account access.
- EuroLLM-22B, an EU-funded project, offers Apache 2.0 weights and partial training data but does not release the full 4-trillion-token corpus.
- Bielik 11B v3.1 from Poland's SpeakLeash foundation includes weights and training scripts but requires Hugging Face account consent for access.
- Luciole, led by LINAGORA, provides weights (Apache 2.0), training data (CC BY-SA 4.0), and code (GPL-3.0), but post-training is almost entirely in English despite supporting 8 languages in the base model.
- Mistral Large 4's license restricts commercial API use without a separate license, while internal server use is free.
- The EU's AI Act requires public summaries of training content for Mistral Large 4, Kolibri-1, and PLLuM, but does not disclose raw data.
Context
Europe's push for open LLMs reflects a strategic effort to reduce reliance on U.S. and Chinese models while fostering local innovation. However, the fragmented approach to licensing and data transparency risks undermining these goals by creating legal and technical barriers.
The EU's AI Act, which mandates transparency in training data summaries, adds another layer of compliance for developers using these models. This aligns with Europe's broader regulatory focus on accountability in AI systems.
Compared to the U.S., where models like Meta's Llama 3 are widely adopted with permissive licenses, European open LLMs often impose stricter conditions—either through gated access, usage policies, or limited data/code releases. This could slow adoption in commercial applications where reproducibility and scalability are critical.
