Soofi S 30B model debuts from German AI consortium

A German research group coordinated by the German AI Association has released the Soofi S 30B model, an open language model trained on Deutsche Telekom’s Industrial AI Cloud in Munich. The consortium says the model leads fully open peers on aggregated German and English benchmarks, a claim now accompanied by a revised report after a training-data contamination issue affected one science benchmark.

Soofi S 30B-A3B has 31.6 billion total parameters but activates about 3.2 billion per token. The consortium adopted Nvidia’s Nemotron 3 Nano architecture without changes, using a mixture-of-experts design that combines Mamba-2 layers with standard attention layers. The practical pitch is serving cost: the model is positioned as a 30B-class system with per-token compute closer to a much smaller model.

What is Soofi S 30B?

Soofi S is an open large language model built by German research institutions and AI companies, with a data mix weighted more heavily toward German than common international recipes. Its design uses only six attention layers that maintain a KV cache, which the consortium says helps keep generation speed steadier as context length increases.

The model was trained from March to May on as many as 512 Nvidia B200 GPUs, totaling about 253,000 GPU-hours, according to the project’s report. Deutsche Telekom’s Munich facility was used for the run. The consortium says the site uses renewable energy, Eisbach canal water for cooling and waste-heat reuse in the Tucherpark area.

Benchmarks came with a contamination correction

The consortium’s version 3.0 pretraining report disclosed that community review found a contamination problem in the QA-base dataset. The dataset was intended to include rephrased training splits from 25 standard benchmarks, but it also contained rephrased questions from GPQA test material, including GPQA Diamond, in English and machine-translated German.

The report attributes the error to GPQA’s Hugging Face labeling: the benchmark has no separate training split, and its test material appears under the default “train” label. After an audit, the team found the same split-label pattern in TruthfulQA, BLiMP and Inverse Scaling, though those benchmarks were not used in the Soofi S evaluation.

The consortium removed GPQA from the evaluation and recalculated results for all 16 comparison models. According to the authors, the ranking order did not change. They also said the model’s GPQA score rose from 32.3 to 43.4 during training, but without a sharp jump that would have made the issue obvious. The team now says it cross-checks training data directly against test questions rather than relying on dataset split names.

German data is the main product decision

The training run used about 27 trillion tokens across three phases. The first phase used roughly 20 trillion tokens from web, code, math and domain-specific text; the second added about 6 trillion tokens from higher-quality sources; the third used long documents of up to one million tokens to extend context.

German text accounted for 7.2 percent of the first phase and 15.3 percent of the second phase, compared with about 5 percent for all non-English languages in Nvidia’s reference recipe, according to the report. The data sources included HPLT German web text, German Commons, German portions of FinePDFs and FineWiki, and the commercially licensed Genios corpus, which contains 193 million newspaper articles from 916 German publications. Machine-translated and synthetic German text also appeared in the mix.

In benchmark comparisons, the consortium says Soofi S led fully open models such as Allen Institute for AI’s OLMo 3 32B and the ETH Zurich and EPFL-backed Apertus 70B on aggregate English and German scores. Reported code results included 73.8 percent on HumanEval, 70.2 on MBPP and 84.2 on the German MBPP variant. The model tied Alibaba’s Qwen3.5 35B-A3B at 61.2 points on INCLUDE-DE, a Germany-specific knowledge benchmark.

The results were not uniform. Soofi S scored 56 on Minerva MATH-DE, behind Qwen3.5 35B-A3B at 76.5 and Gemma 3 27B at 65.6. It also lagged on NaturalQuestions. On one RULER long-context extraction task, its hit rate fell to about 3 percent beyond 32,000 tokens, while the comparable Nemotron model reached 60 to 64 percent. The authors attributed that gap to long-context training data that lacked synthetic extraction tasks.

Licensing and openness remain partly unsettled

The consortium is releasing model weights, selected intermediate checkpoints, training and evaluation code, and a data inventory with token counts, epoch counts and source contributions. It says roughly 152,000 individual evaluation results are available for checking, and that a separate evaluation by partner Ellamind used withheld test data the model had not seen during training.

The project says Soofi S meets the Open Source Initiative’s Open Source AI Definition 1.0. A stricter European open-data proposal would not be met because 1.3 percent of the training mix came from commercially licensed Genios data. The report says about 99 percent of the training mix can be reconstructed independently. The final model license has not been announced.

Participants include Fraunhofer IAIS and IIS, DFKI, TU Darmstadt, the University of Würzburg, L3S Research Center, Berlin University of Applied Sciences, Ellamind and Merantix Momentum. The consortium says instruct and reasoning variants are in beta testing and expected under a permissive license in the coming weeks, while a larger Soofi L model is already in training.

This story draws on original reporting from The Decoder.