Pharia-1

Pharia-1 is a family of open-weight 7-billion-parameter language models from German startup Aleph Alpha, released alongside its PhariaAI enterprise stack. The two variants, Pharia-1-LLM-7B-control and Pharia-1-LLM-7B-control-aligned, are tuned for concise, length-controlled responses in seven European languages with special optimization for English, German, French, and Spanish. Aleph Alpha curated training data to comply with GDPR and the EU AI Act, removing content from millions of websites, and publishes training code, data details, and capability reports.

The aligned variant adds safety training for sensitive topics, and the company openly shares evaluation results including areas where it trails larger models. The weights are released for non-commercial research and educational use, so commercial deployment requires a separate arrangement. Pharia-1 arrived as Aleph Alpha pivoted from frontier-model competition toward enterprise AI tooling, governance, and on-premise deployment, and the models have been adopted by German federal and state administrations.