Jan Pfister

The Soofi-Team, Benedikt Droste, David Fitzek et al.

We present Soofi S 30B-A3B, a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba Transformer foundation model for German and English. Its hybrid design activates only 3B of 30B parameters per token and keeps the inference cache near-constant as context grows, giving it a decisive throughput advantage over dense models for long-context, high-concurrency deployment. Pretrained on roughly 27 trillion tokens with deliberately up-weighted German, Soofi S matches dense 14 to 27B models on aggregate English and German benchmarks while achieving the best code aggregates in both languages among 17 open base models, and outperforms every European sovereign baseline in our comparison, including ones far larger in active parameters. Among fully open models, Soofi S obtains the highest English and German evaluation scores, ahead of Olmo 3 32B and Apertus 70B. Soofi S was built end-to-end on the German Industrial AI Cloud, a sovereign HPC scale AI infrastructure operated by Deutsche Telekom in Munich. Soofi S will be released under highly permissive, open-access terms: weights, selected intermediate checkpoints, full per-source data accounting, hyperparameters, and training and evaluation code. Where source licenses permit, data-construction artifacts are released under permissive licenses; commercially licensed sources are documented with aggregate statistics and exact mixture accounting.

Tobias Koopmann, Jan Pfister, André Markus et al.

Analyzing, understanding, and describing human behavior is advantageous in different settings, such as web browsing or traffic navigation. Understanding human behavior naturally helps to improve and optimize the underlying infrastructure or user interfaces. Typically, human navigation is represented by sequences of transitions between states. Previous work suggests to use hypotheses, representing different intuitions about the navigation to analyze these transitions. To mathematically grasp this setting, first-order Markov chains are used to capture the behavior, consequently allowing to apply different kinds of graph comparisons, but comes with the inherent drawback of losing information about higher-order dependencies within the sequences. To this end, we propose to analyze entire sequences using autoregressive language models, as they are traditionally used to model higher-order dependencies in sequences. We show that our approach can be easily adapted to model different settings introduced in previous work, namely HypTrails, MixedTrails and even SubTrails, while at the same time bringing unique advantages: 1. Modeling higher-order dependencies between state transitions, while 2. being able to identify short comings in proposed hypotheses, and 3. naturally introducing a unified approach to model all settings. To show the expressiveness of our approach, we evaluate our approach on different synthetic datasets and conclude with an exemplary analysis of a real-world dataset, examining the behavior of users who interact with voice assistants.

Julia Wunderle, Anton Ehrmanntraut, Jan Pfister et al.

Encoders remain essential for efficient German NLP and NLU scenarios despite the rise of decoder-only LLMs. This work studies two routes to high-quality German encoders under identical data and training constraints: 1) training from scratch and 2) converting decoders via LLM2Vec. We introduce two resources: ModernGBERT (134M, 1B), fully transparent German encoders in the ModernBERT style, and LLäMmleinVec (120M, 1B, 7B), decoder-to-encoder conversions trained with masked next-token prediction, both undergoing a context extension to 8.192 tokens. Across SuperGLEBer, ModernGBERT 1B sets a new state of the art (avg 0.808), surpassing GBERT Large (+4%) and the seven-times larger converted 7B model (0.787). On German MTEB after supervised fine-tuning, ModernGBERT 1B (0.551) approaches the converted 7B model (0.557). We release all models, checkpoints, datasets, and full training records, and introduce an encoder-adapted QA-NIAH evaluation. All in all, our results provide actionable guidance: when parameter efficiency and latency matter, from-scratch encoders dominate. When a pre-trained decoder exists and compute is a limited, conversion offers an effective alternative. ModernGBERT and LLäMmleinVec, including all code, data and intermediary checkpoints are published under a research-only RAIL license.

Jan Pfister, Julia Wunderle, Andreas Hotho · meta-ai

We create two German-only decoder models, LLäMmlein 120M and 1B, transparently from scratch and publish them, along with the training data, for the German NLP research community to use. The model training involved several key steps, including extensive data preprocessing, the creation of a custom German tokenizer, the training itself, as well as the evaluation of the final models on various benchmarks. Throughout the training process, multiple checkpoints were saved and analyzed using the SuperGLEBer benchmark to monitor the models' learning dynamics. Compared to state-of-the-art models on the SuperGLEBer benchmark, both LLäMmlein models performed competitively, consistently matching or surpassing models with similar parameter sizes. The results show that the models' quality scales with size as expected, but performance improvements on some tasks plateaued early, offering valuable insights into resource allocation for future model development.