German AI Consortium Launches Soofi S: Efficient Open-Source 30B Model That Outperforms Larger Models
Key Takeaways
- ▸Soofi S achieves state-of-the-art performance among fully open models while using only 3.2B active parameters per token, demonstrating the effectiveness of mixture-of-experts optimization for efficiency
- ▸The model excels in both English and German languages, with training deliberately weighted toward German, addressing a gap in high-quality open-source models for European languages
- ▸Efficient long-context processing with minimal throughput degradation enables practical deployment for document processing and extended interactions at scale
Summary
A German research consortium coordinated by the KI Bundesverband has released Soofi S 30B-A3B, an open-source language model trained entirely on Deutsche Telekom's AI cloud infrastructure. The model features a resource-efficient mixture-of-experts (MoE) architecture that activates only 3.2 billion of its 31.6 billion parameters per token, significantly reducing computational requirements while maintaining competitive performance. Trained on 27 trillion tokens with deliberate emphasis on German language data, Soofi S achieves the highest scores on English and German benchmarks among fully open models, surpassing previous leaders like OLMo 3 32B and Apertus 70B.
The model combines Mamba-2 layers with standard attention mechanisms in a hybrid architecture designed for efficient long-context processing. Only 6 of the 52 layers maintain a KV cache, enabling consistent throughput across varying input lengths. At a context length of 40,000 tokens with 32 parallel requests, Soofi S generates approximately eight times more tokens per second per GPU compared to dense 14-24B parameter models, with throughput remaining nearly flat from 4,000 to 256,000 tokens—similar behavior only seen in Alibaba's Qwen3.5 35B-A3B.
The release sparked technical debate around classical scaling laws: critics argued Soofi S was heavily 'overtrained' relative to Google DeepMind's Chinchilla scaling principles (20 tokens per parameter). Project technical lead Michael Fromm countered that traditional scaling laws don't apply to mixture-of-experts architectures, where individual experts benefit from repeated exposure to high-quality data—a defense supported by NVIDIA's own training approach using up to 25 trillion tokens.
- Community-driven development by a German research consortium demonstrates growing European AI research capability independent of US-based labs, backed by Deutsche Telekom infrastructure
Editorial Opinion
Soofi S represents a critical inflection point in open-source LLM development: competing through intelligent architecture rather than brute-force parameter scaling. The model's deliberate focus on German language data is particularly significant for European AI sovereignty, proving that specialized open-source models can outperform generic alternatives through thoughtful training curation. The ongoing debate over scaling laws and mixture-of-experts highlights how the AI community is still refining our understanding of compute-optimality—suggesting the next generation of efficient models may require fundamentally different principles than dense transformers.
