German AI Consortium Releases Soofi S Open Model That Leads English and German Benchmarks

German AI consortium releases Soofi S, an open 30B parameter model that outperforms larger competitors on both English and German benchmarks.

The release comes from a research alliance coordinated by the German AI association. The model uses a MoE design mixing Mamba-2 layers with standard attention. It activates roughly 3 billion parameters per token despite a total size of 31.6 billion.

Soofi S was trained entirely on the German Telecom Munich Industrial AI Cloud. German language share in training data rose from 7.2 percent in the first stage to 15.3 percent in the second stage.

Performance across languages stands out

Soofi S records the highest combined English and German scores among fully open models. It surpasses OLMo 3 32B and Apertus 70B on aggregate benchmarks.

On HumanEval the model reaches 73.8 percent. MBPP scores 70.2 percent. The German version of MBPP reaches 84.2 percent. These results hold even though the model stays smaller than several competitors.

Architecture combines efficiency and scale

The hybrid MoE setup activates only about 3 billion parameters per token. This keeps inference costs low while preserving capacity for complex tasks.

Context length extends to one million tokens. At 40 thousand token length the generation throughput runs about eight times faster than dense models of similar total parameter count. The design pairs Mamba-2 state-space layers with attention layers to balance speed and recall.

Open weights allow direct inspection

The consortium published the weights under an open license. Researchers can download the checkpoint and run local experiments without corporate gatekeepers.

This availability contrasts with several recent large releases that keep weights behind API access only. Independent groups can now measure German-language capability on their own test sets.

Training data choices reflect regional priorities

Raising German content from 7.2 percent to 15.3 percent improved performance on German MBPP without hurting English results. The targeted data mix shows how deliberate language weighting can lift bilingual accuracy.

The model still trails some closed systems on certain English reasoning suites, yet the gap narrows compared with earlier open releases of similar size.

Efficiency claims face real workload tests

Not every production task benefits from the reported eightfold throughput gain. Long-context scenarios with heavy attention patterns may reduce the observed speed advantage.

Users will need to measure token generation on their specific hardware and context lengths before assuming uniform gains. Early adopters should publish their own benchmarks to confirm the lab numbers.

Next signals to monitor

Watch for fine-tuned variants released by community labs within the next three months. Check whether German academic benchmarks continue to improve relative to English ones. Observe whether other national consortia announce comparable bilingual releases that match or exceed these scores.

These three data points will show whether the Soofi S approach scales beyond a single training run or remains a one-off regional success.