A German AI consortium has released Soofi, an open-source 30-billion-parameter language model that tops benchmarks in both English and German.
Developed by a group of German research institutions and companies, Soofi outperforms comparable models like Llama 2 and Mistral on standard evaluations. The model is designed to bridge the gap between English-dominant AI and high-quality German-language performance.
The consortium made the model fully open-source, releasing weights, training code, and evaluation data under a permissive license.
How Soofi Achieves Top Benchmark Scores
Soofi’s architecture is based on a dense transformer with 30 billion parameters. Training used a carefully curated dataset that balanced English and German content.
The model achieved state-of-the-art results on multiple benchmarks including MMLU (English) and German-specific evaluations like the GermanBench and GermEval.
Performance gains came from:
- Higher-quality German training data – The consortium sourced native German texts from books, news, and academic papers, reducing translation artifacts.
- Bilingual tokenization – A specialized tokenizer ensures efficient encoding of German compound words and Umlauts.
- Extended training time – The model was trained for longer than typical 30B models, using more compute to converge on both languages.
“Soofi shows that open models can compete with proprietary systems in non-English languages,” said one of the lead researchers.
Why Open-Source Matters for German AI
The consortium explicitly built Soofi to reduce reliance on U.S.-based AI providers. Many German companies and public institutions cannot use cloud-based models due to data privacy regulations.
Open-source access lets organizations run Soofi on their own infrastructure. The model is already available on Hugging Face and can be downloaded and fine-tuned for specific use cases.
Key advantages for German users:
- Full data control – No data leaves the local environment, complying with GDPR and other regulations.
- Customizable – Users can fine-tune Soofi for legal, medical, or technical German without sharing data.
- Transparency – Training data and methods are published, allowing audits for bias and safety.
Benchmark Results in Detail
On English benchmarks, Soofi matched or exceeded Llama 2 70B on several tasks despite having fewer parameters. On German benchmarks, it outperformed every open model at its size class.
For example, on the German National Competence Test (GNCT) for reading comprehension, Soofi scored 87%—beating the next best open model by 12 percentage points.
On translation tasks between English and German, it achieved BLEU scores comparable to commercial systems from DeepL and Google Translate.
Availability and License
Soofi is released under the Apache 2.0 license. The consortium provides pre-trained weights and a full training pipeline on GitHub.
Researchers recommend using the model with 4-bit quantization for consumer GPUs. A 8-bit version runs on a single A100 80GB.
The consortium includes members from TU Munich, DFKI, and several Berlin-based AI startups. They plan to release a larger version later this year.
“We want to set a standard for sovereign AI in Europe. Soofi is the first step.”
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.