Soofi S: Sovereign Doesn’t Mean Superior

Soofi S 30B-A3B achieves the highest scores on German benchmarks among current European models. Its hybrid MoE-Mamba architecture delivers higher decode throughput at 40K context than pure attention models of the same size class. For an infrastructure team, the decision to deploy ultimately comes down to one pragmatic question: Do these strengths outweigh the beta status and the lack of operational processes?

Key Points in Brief

Related:Almost at top level, trained more cheaply  /  Sovereign Cloud doesn’t end at the server location

Performance Comparison with Qwen and European Models

Soofi scores 79.1 points on the German aggregates, ahead of Nemotron at 74.9. The gap widens on German code. MBPP-DE reaches 84.2, marking the best value among the compared models there. These figures are manufacturer-provided and based on training data with strong German weighting.

Qwen3.5 35B outperforms Soofi on the English aggregate, 74.6 to 70.1. The deficit is equally clear on reasoning benchmarks such as GPQA and BBH. Workloads with a high share of English or abstract reasoning tasks therefore show a noticeable disadvantage versus the Chinese model. The table lists the direct comparison values across all test fields.

The European models Teuken-7B, Apertus, EuroLLM and Salamandra lag further behind in the aggregates. Teuken-7B sits at the bottom left of the chart. Aleph Alpha was acquired by Cohere in April 2026. According to the benchmarks presented, Soofi is currently the strongest open European model of this scale that is not under non-European corporate ownership.

What the Hybrid MoE-Mamba Design Means in Practice

Soofi uses a total of 30 billion parameters but activates only around three billion per token. The Mamba component replaces parts of the classic attention mechanism and scales linearly with context length. This keeps decode speed in tokens per second per GPU high even at 40K context. Full attention models struggle here with the growing KV cache, which consumes memory bandwidth and reduces possible batch sizes. The Mamba component instead maintains a constant state and does not require a KV cache.

For an infrastructure team, this translates into lower latency with long documents, extended chat histories, or retrieval contexts. A key distinction matters in practice and can prove costly if overlooked: the active parameter share reduces compute per token, not weight storage. All 30 billion parameters remain in VRAM; savings occur in FLOPs and activation memory. An additional cost arises from the MoE routing logic. It introduces overhead for batching, load balancing, and scheduling expert activations, plus communication overhead in multi-GPU operation. The diagram positions Soofi according to capability and measured throughput.

It shows the model in the top right, at high capability and the highest measured throughput. No European model has been in this field until now. The position results from the combination of MoE selectivity and Mamba efficiency with long context.

Resource Requirements and What Is Realistic Today

According to the manufacturer, training consumed approximately 253,000 B200 GPU hours on Deutsche Telekom’s Industrial AI Cloud in Munich. Around 27,000 billion training tokens with a strong German weighting were used. The license is permissive, but access is currently restricted to a closed beta.

During inference, the computational effort per token is therefore closer to that of a model with three billion active parameters. A team saves compute time but must still allocate memory for the weights of a 30-billion model. The beta lacks stable versions, documented update paths, and clear information on compatibility with common inference frameworks such as vLLM or TensorRT-LLM.

What is realistically possible today are internal feasibility studies, long-context experiments, and initial fine-tuning on proprietary data. For production systems with defined availability and latency SLAs, the current state is usually insufficient. Integration into existing MLOps pipelines requires additional in-house development and validation work.

The Efficiency Debate on Training Effort

Critics in the expert community point to roughly 80 percent architectural and mixture overlap with Nvidia’s Nemotron 3 Nano. In their view, continual pretraining on the existing model would have required far less compute. The 253.000 B200 GPU hours therefore appear excessive from this perspective. Some observers also regard the self-defined capability index metric as overstated.

The team led by Michael Fromm pushes back: the objective was never the lowest possible eval score. A model trained from scratch with its own data pipeline offers better technical reproducibility and more targeted further development than simply building on an external checkpoint.

The technical trade-off is straightforward. Teams that require full control over training data and weighting accept the higher cost. Those looking for a strong baseline with minimal additional effort tend to choose continual pretraining or an off-the-shelf model.

Who is Responsible for Operations

In production, a team manages patches, security updates, and incident response itself. There is no commercial provider that assumes responsibility for SLAs, monitoring, or fixed response times. Any outage or security vulnerability falls entirely on the operator.

The training work originates from a research consortium. Further development depends on the continuation of this project. Anyone deploying Soofi creates a dependency on a development process that is not primarily designed for operational stability. This increases the effort required for in-house observability, rollback strategies, and emergency procedures. This exact point remains the toughest open question when moving from benchmark to real-world operation.

Frequently Asked Questions

Source cover image: AI-generated (July 2026)