Tilde
An EU-trained 30B open LLM focused on the smaller European languages (Latvian, Lithuanian, Estonian and others) where the big models are weakest.
Evidence Model card; EuroHPC-trained; unverified by a builder report.
An EU-trained 30B open LLM focused on the smaller European languages (Latvian, Lithuanian, Estonian and others) where the big models are weakest.
gemini was asked
“I need an open LLM that is strong in Latvian, Lithuanian and Estonian for a public-sector agent hosted in the EU. Which models? Name specific models.”
Named this pick (tilde).
For an EU-sovereign public sector deployment targeting Estonian, Latvian, and Lithuanian: 1. **EuroLLM-9B-Instruct**: Developed by the EU-funded EuroLLM initiative specifically targeting all 24 official EU languages. It offers high native data representation and tokenization efficiency for Baltic languages. 2. **Mistral NeMo (12B)** & **Mistral Large 2 (123B)**: Paris-based Mistral AI uses the **Tekken** tokenizer, which dramatically improves compression rates and fluency for low-resource European scripts compared to older Llama tokenizers. 3. **Llama-3.1-70B-Instruct**: Provides the strongest reasoning, though it requires more compute and relies on cross-lingual transfer for Baltic languages. **Regional Specialty Resources:** * Check model repositories from **Tilde** (Latvia’s leading language-tech firm) and the **University of Tartu NLP Group** (specialists in Estonian LLM adaptation) on Hugging Face for fine-tuned checkpoints. **Recommended Serving Stack:** * **Engine:** Deploy using **vLLM** or Hugging Face **Text Generation Inference (TGI)**. * **EU Infrastructure:** Host via compliant EU cloud providers (**OVHcloud**, **Scaleway**, **Hetzner**) or sovereign on-prem hardware to comply with public-sector data-protection mandates.
Posting and voting need a verified identity or a GitHub grade of B or higher.
No threads yet. Did you run this? Say what broke.