A German research consortium has released Soofi S, an open-source language model trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. The project, coordinated by the German AI Association (KI Bundesverband) and funded by the German Federal Ministry for Economic Affairs and Energy, utilizes a hybrid Mamba-Transformer architecture. The model contains 31.6 billion total parameters but activates only 3.2 billion per token, which reduces compute costs to levels comparable with a 3B model. The training process occurred between March and May 2026, consuming approximately 253,000 GPU-hours on 512 Nvidia B200 units.

Soofi S demonstrates high efficiency in long-context processing by limiting the KV cache (memory storing previous tokens for attention) to only 6 of its 52 layers. At a context length of 40,000 tokens with 32 parallel requests, the model generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. Throughput remains stable across context lengths from 4,000 to 256,000 tokens. The model's training involved approximately 27 trillion tokens across three phases, with a significant emphasis on German language data. The German-language share of the training mix rose from 7.2 percent in the first phase to 15.3 percent in the second.

The model outperforms other fully open-source alternatives like OLMo 3 32B and Apertus 70B on aggregate German and English benchmarks. Specifically, Soofi S scored 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant. On the Germany-specific knowledge test INCLUDE-DE, it tied for first place at 61.2 points. Despite these results, it scored 56 points on Minerva MATH-DE and struggles with word extraction tasks in the RULER benchmark beyond 32,000 tokens. The consortium published extensive documentation of the training data and code, aiming to align with the Open Source Initiative's Open Source AI Definition 1.0, though the inclusion of commercially licensed Genios data prevents the model from meeting a stricter "fully free" data definition.

A German research consortium has released Soofi S, an open-source language model trained entirely on Deutsche Telekom's Industrial AI Cloud in Munich. The project, coordinated by the German AI Association (KI Bundesverband) and funded by the German Federal Ministry for Economic Affairs and Energy, utilizes a hybrid Mamba-Transformer architecture. The model contains 31.6 billion total parameters but activates only 3.2 billion per token, which reduces compute costs to levels comparable with a 3B model. The training process occurred between March and May 2026, consuming approximately 253,000 GPU-hours on 512 Nvidia B200 units.

Soofi S demonstrates high efficiency in long-context processing by limiting the KV cache (memory storing previous tokens for attention) to only 6 of its 52 layers. At a context length of 40,000 tokens with 32 parallel requests, the model generates roughly eight times more tokens per second per GPU than dense models in the 14 to 24 billion parameter range. Throughput remains stable across context lengths from 4,000 to 256,000 tokens. The model's training involved approximately 27 trillion tokens across three phases, with a significant emphasis on German language data. The German-language share of the training mix rose from 7.2 percent in the first phase to 15.3 percent in the second.

The model outperforms other fully open-source alternatives like OLMo 3 32B and Apertus 70B on aggregate German and English benchmarks. Specifically, Soofi S scored 73.8 percent on HumanEval, 70.2 on MBPP, and 84.2 on the German MBPP variant. On the Germany-specific knowledge test INCLUDE-DE, it tied for first place at 61.2 points. Despite these results, it scored 56 points on Minerva MATH-DE and struggles with word extraction tasks in the RULER benchmark beyond 32,000 tokens. The consortium published extensive documentation of the training data and code, aiming to align with the Open Source Initiative's Open Source AI Definition 1.0, though the inclusion of commercially licensed Genios data prevents the model from meeting a stricter "fully free" data definition.